def get_value(self, plan, penalty_coeff=1e0):
model = grb.Model()
model.params.OutputFlag = 0
self._prob = Prob(model)
# param_to_ll_old = self._param_to_ll.copy()
self._spawn_parameter_to_ll_mapping(model, plan)
model.update()
self._bexpr_to_pred = {}
obj_bexprs = self._get_trajopt_obj(plan)
self._add_obj_bexprs(obj_bexprs)
self._add_all_timesteps_of_actions(plan,
priority=1,
add_nonlin=True,
verbose=False)
return self._prob.get_value(penalty_coeff)
def _solve_opt_prob(self, plan, priority, callback=None, active_ts=None,
verbose=False, resample=True):
## active_ts is the inclusive timesteps to include
## in the optimization
if active_ts==None:
active_ts = (0, plan.horizon-1)
model = grb.Model()
model.params.OutputFlag = 0
self._prob = Prob(model, callback=callback)
# param_to_ll_old = self._param_to_ll.copy()
self._spawn_parameter_to_ll_mapping(model, plan, active_ts)
model.update()
self._bexpr_to_pred = {}
if priority == -1:
obj_bexprs = self._get_trajopt_obj(plan, active_ts)
self._add_obj_bexprs(obj_bexprs)
self._add_first_and_last_timesteps_of_actions(plan, priority=-1, active_ts=active_ts, verbose=verbose)
tol = 1e-2
elif priority == 0:
## this should only get called with a full plan for now
assert active_ts == (0, plan.horizon-1)
obj_bexprs = []
if resample:
failed_preds = plan.get_failed_preds()
## this is an objective that places
## a high value on matching the resampled values
obj_bexprs.extend(self._resample(plan, failed_preds))
## solve an optimization movement primitive to
## transfer current trajectories
obj_bexprs.extend(self._get_transfer_obj(plan, self.transfer_norm))
self._add_obj_bexprs(obj_bexprs)
# self._add_first_and_last_timesteps_of_actions(
# plan, priority=0, add_nonlin=False)
self._add_first_and_last_timesteps_of_actions(
plan, priority=-1, add_nonlin=True, verbose=verbose)
self._add_all_timesteps_of_actions(
plan, priority=0, add_nonlin=False, verbose=verbose)
tol = 1e-2
elif priority == 1:
obj_bexprs = self._get_trajopt_obj(plan, active_ts)
self._add_obj_bexprs(obj_bexprs)
self._add_all_timesteps_of_actions(plan, priority=1, add_nonlin=True, active_ts=active_ts, verbose=verbose)
tol=1e-4
solv = Solver()
solv.initial_penalty_coeff = self.init_penalty_coeff
success = solv.solve(self._prob, method='penalty_sqp', tol=tol, verbose=verbose)
self._update_ll_params()
self._failed_groups = self._prob.nonconverged_groups
return success
def get_value(self, plan, penalty_coeff=1e0):
model = grb.Model()
model.params.OutputFlag = 0
self._prob = Prob(model)
# param_to_ll_old = self._param_to_ll.copy()
self._spawn_parameter_to_ll_mapping(model, plan)
model.update()
self._bexpr_to_pred = {}
obj_bexprs = self._get_trajopt_obj(plan)
self._add_obj_bexprs(obj_bexprs)
self._add_all_timesteps_of_actions(plan, priority=1, add_nonlin=True, verbose=False)
return self._prob.get_value(penalty_coeff)
def monitor_update(self, plan, update_values, callback=None, n_resamples=5, active_ts=None, verbose=False):
print 'Resolving after environment update...\n'
if callback is not None: viewer = callback()
if active_ts==None:
active_ts = (0, plan.horizon-1)
plan.save_free_attrs()
model = grb.Model()
model.params.OutputFlag = 0
self._prob = Prob(model, callback=callback)
# _free_attrs is paied attentioned in here
self._spawn_parameter_to_ll_mapping(model, plan, active_ts)
model.update()
self._bexpr_to_pred = {}
tol = 1e-3
obj_bexprs = []
rs_obj = self._update(plan, update_values)
obj_bexprs.extend(self._get_transfer_obj(plan, self.transfer_norm))
self._add_all_timesteps_of_actions(plan, priority=MAX_PRIORITY,
add_nonlin=True, active_ts= active_ts, verbose=verbose)
obj_bexprs.extend(rs_obj)
self._add_obj_bexprs(obj_bexprs)
initial_trust_region_size = 1e-2
solv = Solver()
solv.initial_trust_region_size = initial_trust_region_size
solv.initial_penalty_coeff = self.init_penalty_coeff
solv.max_merit_coeff_increases = self.max_merit_coeff_increases
success = solv.solve(self._prob, method='penalty_sqp', tol=tol, verbose=verbose)
self._update_ll_params()
if DEBUG: assert not plan.has_nan(active_ts)
plan.restore_free_attrs()
self.reset_variable()
print "monitor_update\n"
return success
class NAMOSolver(LLSolver):
def __init__(self, early_converge=True, transfer_norm='min-vel'):
self.transfer_coeff = 1e1
self.rs_coeff = 1e6
self.init_penalty_coeff = 1e2
self.child_solver = None
self._param_to_ll = {}
self._failed_groups = []
self.early_converge=early_converge
self.transfer_norm = transfer_norm
def backtrack_solve(self, plan, callback=None, verbose=False):
plan.save_free_attrs()
success = self._backtrack_solve(plan, callback, anum=0, verbose=verbose)
plan.restore_free_attrs()
return success
def _backtrack_solve(self, plan, callback=None, anum=0, verbose=False):
if anum > len(plan.actions) - 1:
return True
a = plan.actions[anum]
active_ts = a.active_timesteps
inits = {}
if a.name == 'moveto':
## find possible values for the final pose
rs_param = a.params[2]
elif a.name == 'movetoholding':
## find possible values for the final pose
rs_param = a.params[2]
elif a.name == 'grasp':
## sample the grasp/grasp_pose
rs_param = a.params[4]
elif a.name == 'putdown':
## sample the end pose
rs_param = a.params[4]
else:
raise NotImplemented
def recursive_solve():
## don't optimize over any params that are already set
old_params_free = {}
for p in plan.params.itervalues():
if p.is_symbol():
if p not in a.params: continue
old_params_free[p] = p._free_attrs['value'].copy()
p._free_attrs['value'][:] = 0
else:
old_params_free[p] = p._free_attrs['pose'][:, active_ts[1]].copy()
p._free_attrs['pose'][:, active_ts[1]] = 0
self.child_solver = NAMOSolver()
if self.child_solver._backtrack_solve(plan, callback=callback, anum=anum+1, verbose=verbose):
return True
## reset free_attrs
for p in a.params:
if p.is_symbol():
if p not in a.params: continue
p._free_attrs['value'] = old_params_free[p]
else:
p._free_attrs['pose'][:, active_ts[1]] = old_params_free[p]
return False
if not np.all(rs_param._free_attrs['value']):
## this parameter is fixed
if callback is not None:
callback_a = lambda: callback(a)
else:
callback_a = None
self.child_solver = NAMOSolver()
success = self.child_solver.solve(plan, callback=callback_a, n_resamples=0,
active_ts = active_ts, verbose=verbose, force_init=True)
if not success:
## if planning fails we're done
return False
## no other options, so just return here
return recursive_solve()
## so that this won't be optimized over
rs_free = rs_param._free_attrs['value'].copy()
rs_param._free_attrs['value'][:] = 0
targets = plan.get_param('InContact', 2, {1:rs_param}, negated=False)
if len(targets) > 1:
import pdb; pdb.set_trace()
if callback is not None:
callback_a = lambda: callback(a)
else:
callback_a = None
robot_poses = []
if len(targets) == 0 or np.all(targets[0]._free_attrs['value']):
## sample 4 possible poses
coords = list(itertools.product(range(WIDTH), range(HEIGHT)))
random.shuffle(coords)
robot_poses = [np.array(x)[:, None] for x in coords[:4]]
elif np.any(targets[0]._free_attrs['value']):
## there shouldn't be only some free_attrs set
raise NotImplementedError
else:
grasp_dirs = [np.array([0, -1]),
np.array([1, 0]),
np.array([0, 1]),
np.array([-1, 0])]
grasp_len = plan.params['pr2'].geom.radius + targets[0].geom.radius - dsafe
for g_dir in grasp_dirs:
grasp = (g_dir*grasp_len).reshape((2, 1))
robot_poses.append(targets[0].value + grasp)
for rp in robot_poses:
rs_param.value = rp
success = False
self.child_solver = NAMOSolver()
success = self.child_solver.solve(plan, callback=callback_a, n_resamples=0,
active_ts = active_ts, verbose=verbose,
force_init=True)
if success:
if recursive_solve():
break
else:
success = False
rs_param._free_attrs['value'] = rs_free
return success
def solve(self, plan, callback=None, n_resamples=5, active_ts=None, verbose=False, force_init=False):
success = False
if force_init or not plan.initialized:
## solve at priority -1 to get an initial value for the parameters
self._solve_opt_prob(plan, priority=-1, callback=callback, active_ts=active_ts, verbose=verbose)
plan.initialized=True
success = self._solve_opt_prob(plan, priority=1, callback=callback, active_ts=active_ts, verbose=verbose)
success = plan.satisfied(active_ts)
if success:
return success
for _ in range(n_resamples):
## refinement loop
## priority 0 resamples the first failed predicate in the plan
## and then solves a transfer optimization that only includes linear constraints
self._solve_opt_prob(plan, priority=0, callback=callback, active_ts=active_ts, verbose=verbose)
success = self._solve_opt_prob(plan, priority=1, callback=callback, active_ts=active_ts, verbose=verbose)
success = plan.satisfied(active_ts)
if success:
return success
return success
# @profile
def _solve_opt_prob(self, plan, priority, callback=None, active_ts=None,
verbose=False, resample=True):
## active_ts is the inclusive timesteps to include
## in the optimization
if active_ts==None:
active_ts = (0, plan.horizon-1)
model = grb.Model()
model.params.OutputFlag = 0
self._prob = Prob(model, callback=callback)
# param_to_ll_old = self._param_to_ll.copy()
self._spawn_parameter_to_ll_mapping(model, plan, active_ts)
model.update()
self._bexpr_to_pred = {}
if priority == -1:
obj_bexprs = self._get_trajopt_obj(plan, active_ts)
self._add_obj_bexprs(obj_bexprs)
self._add_first_and_last_timesteps_of_actions(plan, priority=-1, active_ts=active_ts, verbose=verbose)
tol = 1e-2
elif priority == 0:
## this should only get called with a full plan for now
assert active_ts == (0, plan.horizon-1)
obj_bexprs = []
if resample:
failed_preds = plan.get_failed_preds()
## this is an objective that places
## a high value on matching the resampled values
obj_bexprs.extend(self._resample(plan, failed_preds))
## solve an optimization movement primitive to
## transfer current trajectories
obj_bexprs.extend(self._get_transfer_obj(plan, self.transfer_norm))
self._add_obj_bexprs(obj_bexprs)
# self._add_first_and_last_timesteps_of_actions(
# plan, priority=0, add_nonlin=False)
self._add_first_and_last_timesteps_of_actions(
plan, priority=-1, add_nonlin=True, verbose=verbose)
self._add_all_timesteps_of_actions(
plan, priority=0, add_nonlin=False, verbose=verbose)
tol = 1e-2
elif priority == 1:
obj_bexprs = self._get_trajopt_obj(plan, active_ts)
self._add_obj_bexprs(obj_bexprs)
self._add_all_timesteps_of_actions(plan, priority=1, add_nonlin=True, active_ts=active_ts, verbose=verbose)
tol=1e-4
solv = Solver()
solv.initial_penalty_coeff = self.init_penalty_coeff
success = solv.solve(self._prob, method='penalty_sqp', tol=tol, verbose=verbose)
self._update_ll_params()
self._failed_groups = self._prob.nonconverged_groups
return success
def get_value(self, plan, penalty_coeff=1e0):
model = grb.Model()
model.params.OutputFlag = 0
self._prob = Prob(model)
# param_to_ll_old = self._param_to_ll.copy()
self._spawn_parameter_to_ll_mapping(model, plan)
model.update()
self._bexpr_to_pred = {}
obj_bexprs = self._get_trajopt_obj(plan)
self._add_obj_bexprs(obj_bexprs)
self._add_all_timesteps_of_actions(plan, priority=1, add_nonlin=True, verbose=False)
return self._prob.get_value(penalty_coeff)
def _get_transfer_obj(self, plan, norm):
transfer_objs = []
if norm in ['min-vel', 'l2']:
for param in plan.params.values():
# if param._type in ['Robot', 'Can']:
K = 2
if param.is_symbol():
T = 1
pose = param.value
else:
T = plan.horizon
pose = param.pose
assert (K, T) == pose.shape
KT = K*T
if norm == 'min-vel' and not param.is_symbol():
v = -1 * np.ones((KT - K, 1))
d = np.vstack((np.ones((KT - K, 1)), np.zeros((K, 1))))
# [:,0] allows numpy to see v and d as one-dimensional so
# that numpy will create a diagonal matrix with v and d as a diagonal
P = np.diag(v[:, 0], K) + np.diag(d[:, 0])
Q = np.dot(np.transpose(P), P)
else: ## l2
Q = np.eye(KT)
cur_pose = pose.reshape((KT, 1), order='F')
A = -2*cur_pose.T.dot(Q)
b = cur_pose.T.dot(Q.dot(cur_pose))
# QuadExpr is 0.5*x^Tx + Ax + b
quad_expr = QuadExpr(2*self.transfer_coeff*Q,
self.transfer_coeff*A, self.transfer_coeff*b)
ll_param = self._param_to_ll[param]
if param.is_symbol():
ll_grb_vars = ll_param.value.reshape((KT, 1), order='F')
else:
ll_grb_vars = ll_param.pose.reshape((KT, 1), order='F')
bexpr = BoundExpr(quad_expr, Variable(ll_grb_vars, cur_pose))
transfer_objs.append(bexpr)
elif norm == 'straightline':
return self._get_trajopt_obj(plan)
else:
raise NotImplementedError
return transfer_objs
def _resample(self, plan, preds):
val, attr_inds = None, None
for negated, pred, t in preds:
## returns a vector of new values and an
## attr_inds (OrderedDict) that gives the mapping
## to parameter attributes
if (self.early_converge and self._failed_groups != ['all'] and self._failed_groups != []
and not np.any([param.name in self._failed_groups for param in pred.params])):
## using early converge and one group didn't converge
## but no parameter from this pred in that group
continue
val, attr_inds = pred.resample(negated, t, plan)
## no resample defined for that pred
if val is not None: break
if val is None:
# import pdb; pdb.set_trace()
return []
t_local = t
bexprs = []
i = 0
for p in attr_inds:
## get the ll_param for p and gurobi variables
ll_p = self._param_to_ll[p]
if p.is_symbol(): t_local = 0
n_vals = 0
grb_vars = []
for attr, ind_arr in attr_inds[p]:
n_vals += len(ind_arr)
grb_vars.extend(
list(getattr(ll_p, attr)[ind_arr, t_local]))
for j, grb_var in enumerate(grb_vars):
## create an objective saying stay close to this value
## e(x) = x^2 - 2*val[i+j]*x + val[i+j]^2
Q = np.eye(1)
A = -2*val[i+j]*np.ones((1, 1))
b = np.ones((1, 1))*np.power(val[i+j], 2)
# QuadExpr is 0.5*x^Tx + Ax + b
quad_expr = QuadExpr(2*Q*self.rs_coeff, A*self.rs_coeff, b*self.rs_coeff)
v_arr = np.array([grb_var]).reshape((1, 1), order='F')
init_val = np.ones((1, 1))*val[i+j]
bexpr = BoundExpr(quad_expr,
Variable(v_arr, val[i+j].reshape((1, 1))))
bexprs.append(bexpr)
i += n_vals
return bexprs
def _add_pred_dict(self, pred_dict, effective_timesteps, add_nonlin=True, priority=MAX_PRIORITY, verbose=False):
# verbose=True
## for debugging
ignore_preds = []
priority = np.maximum(priority, 0)
if not pred_dict['hl_info'] == "hl_state":
start, end = pred_dict['active_timesteps']
active_range = range(start, end+1)
if verbose:
print "pred being added: ", pred_dict
print active_range, effective_timesteps
negated = pred_dict['negated']
pred = pred_dict['pred']
if pred.get_type() in ignore_preds:
return
if pred.priority > priority: return
assert isinstance(pred, common_predicates.ExprPredicate)
expr = pred.get_expr(negated)
for t in effective_timesteps:
if t in active_range:
if expr is not None:
if add_nonlin or isinstance(expr.expr, AffExpr):
if verbose:
print "expr being added at time ", t
var = self._spawn_sco_var_for_pred(pred, t)
bexpr = BoundExpr(expr, var)
self._bexpr_to_pred[bexpr] = (negated, pred, t)
groups = ['all']
if self.early_converge:
## this will check for convergence per parameter
## this is good if e.g., a single trajectory quickly
## gets stuck
groups.extend([param.name for param in pred.params])
self._prob.add_cnt_expr(bexpr, groups)
def _add_first_and_last_timesteps_of_actions(self, plan, priority = MAX_PRIORITY, add_nonlin=False, active_ts=None, verbose=False):
if active_ts==None:
active_ts = (0, plan.horizon-1)
for action in plan.actions:
action_start, action_end = action.active_timesteps
## only add an action
if action_start >= active_ts[1] and action_start > active_ts[0]: continue
if action_end < active_ts[0]: continue
for pred_dict in action.preds:
if action_start >= active_ts[0]:
self._add_pred_dict(pred_dict, [action_start],
priority=priority, add_nonlin=add_nonlin, verbose=verbose)
if action_end <= active_ts[1]:
self._add_pred_dict(pred_dict, [action_end],
priority=priority, add_nonlin=add_nonlin, verbose=verbose)
## add all of the linear ineqs
timesteps = range(max(action_start+1, active_ts[0]),
min(action_end, active_ts[1]))
for pred_dict in action.preds:
self._add_pred_dict(pred_dict, timesteps, add_nonlin=False, priority=priority, verbose=verbose)
def _add_all_timesteps_of_actions(self, plan, priority=MAX_PRIORITY, add_nonlin=True, active_ts=None, verbose=False):
if active_ts==None:
active_ts = (0, plan.horizon-1)
for action in plan.actions:
action_start, action_end = action.active_timesteps
if action_start >= active_ts[1] and action_start > active_ts[0]: continue
if action_end < active_ts[0]: continue
timesteps = range(max(action_start, active_ts[0]),
min(action_end, active_ts[1])+1)
for pred_dict in action.preds:
self._add_pred_dict(pred_dict, timesteps, priority=priority, add_nonlin=add_nonlin, verbose=verbose)
def _update_ll_params(self):
for ll_param in self._param_to_ll.values():
ll_param.update_param()
if self.child_solver:
self.child_solver._update_ll_params()
def _spawn_parameter_to_ll_mapping(self, model, plan, active_ts=None):
if active_ts == None:
active_ts=(0, plan.horizon-1)
horizon = active_ts[1] - active_ts[0] + 1
self._param_to_ll = {}
self.ll_start = active_ts[0]
for param in plan.params.values():
ll_param = LLParam(model, param, horizon, active_ts)
ll_param.create_grb_vars()
self._param_to_ll[param] = ll_param
model.update()
for ll_param in self._param_to_ll.values():
ll_param.batch_add_cnts()
def _add_obj_bexprs(self, obj_bexprs):
for bexpr in obj_bexprs:
self._prob.add_obj_expr(bexpr)
def _get_trajopt_obj(self, plan, active_ts=None):
if active_ts == None:
active_ts = (0, plan.horizon-1)
start, end = active_ts
traj_objs = []
for param in plan.params.values():
if param not in self._param_to_ll:
continue
if param._type in ['Robot', 'Can']:
T = end - start + 1
K = 2
pose = param.pose
KT = K*T
v = -1 * np.ones((KT - K, 1))
d = np.vstack((np.ones((KT - K, 1)), np.zeros((K, 1))))
# [:,0] allows numpy to see v and d as one-dimensional so
# that numpy will create a diagonal matrix with v and d as a diagonal
P = np.diag(v[:, 0], K) + np.diag(d[:, 0])
Q = np.dot(np.transpose(P), P)
Q *= TRAJOPT_COEFF
quad_expr = QuadExpr(Q, np.zeros((1,KT)), np.zeros((1,1)))
robot_ll = self._param_to_ll[param]
robot_ll_grb_vars = robot_ll.pose.reshape((KT, 1), order='F')
init_val = param.pose[:, start:end+1].reshape((KT, 1), order='F')
bexpr = BoundExpr(quad_expr, Variable(robot_ll_grb_vars, init_val))
traj_objs.append(bexpr)
return traj_objs
class DrivingSolver(LLSolver):
def __init__(self, early_converge=False, transfer_norm='min-vel'):
# To avoid numerical difficulties during optimization, try keep
# range of coefficeint within 1e9
# (largest_coefficient/smallest_coefficient < 1e9)
self.transfer_coeff = 1e0
self.rs_coeff = 5e1
self.trajopt_coeff = 1e0
self.initial_trust_region_size = 1e-2
self.init_penalty_coeff = 4e3
self.smooth_penalty_coeff = 7e4
self.max_merit_coeff_increases = 5
self._param_to_ll = {}
self.early_converge=early_converge
self.child_solver = None
self.solve_priorities = [0, 1, 2, 3]
self.transfer_norm = transfer_norm
self.grb_init_mapping = {}
self.var_list = []
self._grb_to_var_ind = {}
self.tol = 1e-3
def _solve_helper(self, plan, callback, active_ts, verbose):
# certain constraints should be solved first
success = False
for priority in self.solve_priorities:
success = self._solve_opt_prob(plan, priority=priority,
callback=callback, active_ts=active_ts,
verbose=verbose)
return success
def backtrack_solve(self, plan, callback=None, verbose=False, n_resamples=5):
plan.save_free_attrs()
success = self._backtrack_solve(plan, callback, anum=0, verbose=verbose, n_resamples=n_resamples)
# plan.restore_free_attrs()
return success
#@profile
def _backtrack_solve(self, plan, callback=None, anum=0, verbose=False, amax = None, n_resamples=5):
# if anum == 2:
# import ipdb; ipdb.set_trace()
if amax is None:
amax = len(plan.actions) - 1
if anum > amax:
return True
a = plan.actions[anum]
print "backtracking Solve on {}".format(a.name)
active_ts = a.active_timesteps
inits = {}
if a.name == 'drive_down_road':
rs_param = a.params[3]
else:
raise NotImplemented
def recursive_solve():
# import ipdb; ipdb.set_trace()
## don't optimize over any params that are already set
old_params_free = {}
for p in plan.params.itervalues():
if p.is_symbol():
if p not in a.params: continue
old_params_free[p] = p._free_attrs
p._free_attrs = {}
for attr in old_params_free[p].keys():
p._free_attrs[attr] = np.zeros(old_params_free[p][attr].shape)
else:
p_attrs = {}
old_params_free[p] = p_attrs
for attr in p._free_attrs:
p_attrs[attr] = p._free_attrs[attr][:, active_ts[1]].copy()
p._free_attrs[attr][:, active_ts[1]] = 0
self.child_solver = DrivingSolver()
success = self.child_solver._backtrack_solve(plan, callback=callback, anum=anum+1, verbose=verbose, amax = amax)
# reset free_attrs
for p in plan.params.itervalues():
if p.is_symbol():
if p not in a.params: continue
p._free_attrs = old_params_free[p]
else:
for attr in p._free_attrs:
p._free_attrs[attr][:, active_ts[1]] = old_params_free[p][attr]
return success
# if there is no parameter to resample or some part of rs_param is fixed, then go ahead optimize over this action
if rs_param is None or sum([not np.all(rs_param._free_attrs[attr]) for attr in rs_param._free_attrs.keys() ]):
## this parameter is fixed
if callback is not None:
callback_a = lambda: callback(a)
else:
callback_a = None
self.child_solver = DrivingSolver()
success = self.child_solver.solve(plan, callback=callback_a, n_resamples=n_resamples,
active_ts = active_ts, verbose=verbose, force_init=True)
if not success:
## if planning fails we're done
return False
## no other options, so just return here
return recursive_solve()
## so that this won't be optimized over
rs_free = rs_param._free_attrs
rs_param._free_attrs = {}
for attr in rs_free.keys():
rs_param._free_attrs[attr] = np.zeros(rs_free[attr].shape)
"""
sampler_begin
"""
vehicle_poses = self.obj_pose_suggester(plan, anum, resample_size=3)
if not vehicle_poses:
success = False
# print "Using Random Poses"
# robot_poses = self.random_pose_suggester(plan, anum, resample_size = 5)
"""
sampler end
"""
if callback is not None:
callback_a = lambda: callback(a)
else:
callback_a = None
for rp in vehicle_poses:
for attr, val in rp.iteritems():
setattr(rs_param, attr, val)
success = False
self.child_solver = DrivingSolver()
success = self.child_solver.solve(plan, callback=callback_a, n_resamples=n_resamples,
active_ts = active_ts, verbose=verbose,
force_init=True)
if success:
if recursive_solve():
break
else:
success = False
rs_param._free_attrs = rs_free
return success
#@profile
def random_pose_suggester(self, plan, anum, resample_size=5):
pass
#@profile
def obj_pose_suggester(self, plan, anum, resample_size=20):
vehicle_poses = []
assert anum + 1 <= len(plan.actions)
if anum + 1 < len(plan.actions):
act, next_act = plan.actions[anum], plan.actions[anum+1]
else:
act, next_act = plan.actions[anum], None
vehicle = act.params[0]
start_ts, end_ts = act.active_timesteps
for i in range(resample_size):
if act.name == "drive_down_road":
road = act.params[1]
init_pos = act.params[2]
direction = road.geom.direction
dist = road.geom.length / 2
x = road.geom.x
y = road.geom.y
final_xy = x + dist * np.cos(direction), y + dist * np.sin(direction)
vehicle_poses.append({'xy': final_xy, 'theta': direction, 'vel': 0, 'phi': 0, 'u1': 0, 'u2': 0, 'value': 0})
else:
raise NotImplementedError
if not vehicle_poses:
print "Unable to find IK"
# import ipdb; ipdb.set_trace()
return vehicle_poses
#@profile
def solve(self, plan, callback=None, n_resamples=5, active_ts=None,
verbose=False, force_init=False):
success = False
if callback is not None:
viewer = callback()
if force_init or not plan.initialized:
self._solve_opt_prob(plan, priority=-2, callback=callback,
active_ts=active_ts, verbose=verbose)
# self._solve_opt_prob(plan, priority=-1, callback=callback,
# active_ts=active_ts, verbose=verbose)
plan.initialized=True
if success or len(plan.get_failed_preds(active_ts = active_ts)) == 0:
return True
for priority in self.solve_priorities:
for attempt in range(n_resamples):
## refinement loop
success = self._solve_opt_prob(plan, priority=priority,
callback=callback, active_ts=active_ts, verbose=verbose)
try:
if DEBUG: plan.check_cnt_violation(active_ts = active_ts, priority = priority, tol = 1e-3)
except:
print "error in predicate checking"
if success:
break
self._solve_opt_prob(plan, priority=priority, callback=callback, active_ts=active_ts, verbose=verbose, resample = True)
# if len(plan.get_failed_preds(active_ts=active_ts, tol=1e-3)) > 9:
# break
print "resample attempt: {}".format(attempt)
try:
if DEBUG: plan.check_cnt_violation(active_ts = active_ts, priority = priority, tol = 1e-3)
except:
print "error in predicate checking"
assert not (success and not len(plan.get_failed_preds(active_ts = active_ts, priority = priority, tol = 1e-3)) == 0)
if not success:
return False
return success
#@profile
def _solve_opt_prob(self, plan, priority, callback=None, init=True,
active_ts=None, verbose=False, resample=False, smoothing = False):
if callback is not None: viewer = callback()
self.plan = plan
vehicle = plan.params['vehicle']
if active_ts==None:
active_ts = (0, plan.horizon-1)
plan.save_free_attrs()
model = grb.Model()
model.params.OutputFlag = 0
self._prob = Prob(model, callback=callback)
self._spawn_parameter_to_ll_mapping(model, plan, active_ts)
model.update()
initial_trust_region_size = self.initial_trust_region_size
if resample:
tol = 1e-3
def variable_helper():
error_bin = []
for sco_var in self._prob._vars:
for grb_var, val in zip(sco_var.get_grb_vars(), sco_var.get_value()):
grb_name = grb_var[0].VarName
one, two = grb_name.find('-'), grb_name.find('-(')
param_name = grb_name[1:one]
attr = grb_name[one+1:two]
index = eval(grb_name[two+1:-1])
param = plan.params[param_name]
if not np.allclose(val, getattr(param, attr)[index]):
error_bin.append((grb_name, val, getattr(param, attr)[index]))
if len(error_bin) != 0:
print "something wrong"
if DEBUG: import ipdb; ipdb.set_trace()
"""
When Optimization fails, resample new values for certain timesteps
of the trajectory and solver as initialization
"""
obj_bexprs = []
## this is an objective that places
## a high value on matching the resampled values
failed_preds = plan.get_failed_preds(active_ts = active_ts, priority=priority, tol = tol)
rs_obj = self._resample(plan, failed_preds, sample_all = True)
# import ipdb; ipdb.set_trace()
# _get_transfer_obj returns the expression saying the current trajectory should be close to it's previous trajectory.
# obj_bexprs.extend(self._get_trajopt_obj(plan, active_ts))
obj_bexprs.extend(self._get_transfer_obj(plan, self.transfer_norm))
self._add_all_timesteps_of_actions(plan, priority=priority,
add_nonlin=False, active_ts= active_ts, verbose=verbose)
obj_bexprs.extend(rs_obj)
self._add_obj_bexprs(obj_bexprs)
initial_trust_region_size = 1e3
# import ipdb; ipdb.set_trace()
else:
self._bexpr_to_pred = {}
if priority == -2:
"""
Initialize an linear trajectory while enforceing the linear constraints in the intermediate step.
"""
obj_bexprs = self._get_trajopt_obj(plan, active_ts)
self._add_obj_bexprs(obj_bexprs)
self._add_first_and_last_timesteps_of_actions(plan,
priority=MAX_PRIORITY, active_ts=active_ts, verbose=verbose, add_nonlin=False)
tol = 1e-1
initial_trust_region_size = 1e3
elif priority == -1:
"""
Solve the optimization problem while enforcing every constraints.
"""
obj_bexprs = self._get_trajopt_obj(plan, active_ts)
self._add_obj_bexprs(obj_bexprs)
self._add_first_and_last_timesteps_of_actions(plan,
priority=MAX_PRIORITY, active_ts=active_ts, verbose=verbose,
add_nonlin=True)
tol = 1e-1
elif priority >= 0:
obj_bexprs = self._get_trajopt_obj(plan, active_ts)
self._add_obj_bexprs(obj_bexprs)
self._add_all_timesteps_of_actions(plan, priority=priority, add_nonlin=True,
active_ts=active_ts, verbose=verbose)
tol=1e-3
solv = Solver()
solv.initial_trust_region_size = initial_trust_region_size
if smoothing:
solv.initial_penalty_coeff = self.smooth_penalty_coeff
else:
solv.initial_penalty_coeff = self.init_penalty_coeff
solv.max_merit_coeff_increases = self.max_merit_coeff_increases
success = solv.solve(self._prob, method='penalty_sqp', tol=tol, verbose=verbose)
if priority == MAX_PRIORITY:
success = success or len(plan.get_failed_preds(tol=tol, active_ts=active_ts, priority=priority)) == 0
self._update_ll_params()
if DEBUG: assert not plan.has_nan(active_ts)
if resample:
# During resampling phases, there must be changes added to sampling_trace
if len(plan.sampling_trace) > 0 and 'reward' not in plan.sampling_trace[-1]:
reward = 0
if len(plan.get_failed_preds(active_ts = active_ts, priority=priority)) == 0:
reward = len(plan.actions)
else:
failed_t = plan.get_failed_pred(active_ts=(0,active_ts[1]), priority=priority)[2]
for i in range(len(plan.actions)):
if failed_t > plan.actions[i].active_timesteps[1]:
reward += 1
plan.sampling_trace[-1]['reward'] = reward
##Restore free_attrs values
plan.restore_free_attrs()
self.reset_variable()
print "priority: {}\n".format(priority)
return success
#@profile
def traj_smoother(self, plan, callback=None, n_resamples=5, verbose=False):
# plan.save_free_attrs()
a_num = 0
success = True
while a_num < len(plan.actions) - 1:
act_1 = plan.actions[a_num]
act_2 = plan.actions[a_num+1]
active_ts = (act_1.active_timesteps[0], act_2.active_timesteps[1])
# print active_ts
old_params_free = {}
for p in plan.params.itervalues():
if p.is_symbol():
if p in act_1.params or p in act_2.params: continue
old_params_free[p] = p._free_attrs
p._free_attrs = {}
for attr in old_params_free[p].keys():
p._free_attrs[attr] = np.zeros(old_params_free[p][attr].shape)
else:
p_attrs = {}
old_params_free[p] = p_attrs
for attr in p._free_attrs:
p_attrs[attr] = [p._free_attrs[attr][:, :active_ts[0]].copy(), p._free_attrs[attr][:, active_ts[1]:].copy()]
p._free_attrs[attr][:, active_ts[1]:] = 0
p._free_attrs[attr][:, :active_ts[0]] = 0
success = self._traj_smoother(plan, callback, n_resamples, active_ts, verbose)
# reset free_attrs
for p in plan.params.itervalues():
if p.is_symbol():
if p in act_1.params or p in act_2.params: continue
p._free_attrs = old_params_free[p]
else:
for attr in p._free_attrs:
p._free_attrs[attr][:, :active_ts[0]] = old_params_free[p][attr][0]
p._free_attrs[attr][:, active_ts[1]:] = old_params_free[p][attr][1]
if not success:
return success
print 'Actions: {} and {}'.format(plan.actions[a_num].name, plan.actions[a_num+1].name)
a_num += 1
# try:
# success = self._traj_smoother(plan, callback, n_resamples, active_ts, verbose)
# except:
# print "Error occured during planning, but not catched"
# return False
# plan.restore_free_attrs()
return success
# @profile
def _traj_smoother(self, plan, callback=None, n_resamples=5, active_ts=None, verbose=False):
print "Smoothing Trajectory..."
priority = MAX_PRIORITY
for attempt in range(n_resamples):
# refinement loop
print 'Smoother iteration #: {}\n'.format(attempt)
success = self._solve_opt_prob(plan, priority=priority,
callback=callback, active_ts=active_ts, verbose=verbose, resample = False, smoothing = True)
if success:
break
plan.check_cnt_violation(tol = 1e-3)
self._solve_opt_prob(plan, priority=priority, callback=callback, active_ts=active_ts, verbose=verbose, resample = True, smoothing = True)
return success
#@profile
def _get_transfer_obj(self, plan, norm):
"""
This function returns the expression e(x) = P|x - cur|^2
Which says the optimized trajectory should be close to the
previous trajectory.
Where P is the KT x KT matrix, where Px is the difference of parameter's attributes' current value and parameter's next timestep value
"""
transfer_objs = []
if norm == 'min-vel':
for param in plan.params.values():
for attr_name in param.__dict__.iterkeys():
attr_type = param.get_attr_type(attr_name)
if issubclass(attr_type, Vector):
param_ll = self._param_to_ll[param]
if param.is_symbol():
T = 1
attr_val = getattr(param, attr_name)
else:
T = param_ll._horizon
attr_val = getattr(param, attr_name)[:, param_ll.active_ts[0]:param_ll.active_ts[1]+1]
K = attr_type.dim
# pose = param.pose
if DEBUG: assert (K, T) == attr_val.shape
KT = K*T
v = -1 * np.ones((KT - K, 1))
d = np.vstack((np.ones((KT - K, 1)), np.zeros((K, 1))))
# [:,0] allows numpy to see v and d as one-dimensional so
# that numpy will create a diagonal matrix with v and d as a diagonal
P = np.diag(v[:, 0], K) + np.diag(d[:, 0])
# P = np.eye(KT)
Q = np.dot(np.transpose(P), P) if not param.is_symbol() else np.eye(KT)
cur_val = attr_val.reshape((KT, 1), order='F')
A = -2*cur_val.T.dot(Q)
b = cur_val.T.dot(Q.dot(cur_val))
transfer_coeff = self.transfer_coeff/float(plan.horizon)
# QuadExpr is 0.5*x^Tx + Ax + b
quad_expr = QuadExpr(2*transfer_coeff*Q,
transfer_coeff*A, transfer_coeff*b)
ll_attr_val = getattr(param_ll, attr_name)
param_ll_grb_vars = ll_attr_val.reshape((KT, 1), order='F')
sco_var = self.create_variable(param_ll_grb_vars, cur_val)
bexpr = BoundExpr(quad_expr, sco_var)
transfer_objs.append(bexpr)
else:
raise NotImplemented
return transfer_objs
#@profile
def create_variable(self, grb_vars, init_vals, save=False):
"""
if save is Ture
Update the grb_init_mapping so that each grb_var is mapped to
the right initial values.
Then find the sco variables that includes the grb variables we are updating and change the corresponding initial values inside of it.
if save is False
Iterate the var_list and use the last initial value used for each gurobi, and construct the sco variables
"""
sco_var, grb_val_map, ret_val = None, {}, []
for grb, v in zip(grb_vars.flatten(), init_vals.flatten()):
grb_name = grb.VarName
if save: self.grb_init_mapping[grb_name] = v
grb_val_map[grb_name] = self.grb_init_mapping.get(grb_name, v)
ret_val.append(grb_val_map[grb_name])
if grb_name in self._grb_to_var_ind.keys():
for var, i in self._grb_to_var_ind[grb_name]:
var._value[i] = grb_val_map[grb_name]
if np.all(var._grb_vars is grb_vars):
sco_var = var
if sco_var is None:
sco_var = Variable(grb_vars, np.array(ret_val).reshape((len(ret_val), 1)))
self.var_list.append(sco_var)
for i, grb in enumerate(grb_vars.flatten()):
index_val_list = self._grb_to_var_ind.get(grb.VarName, [])
index_val_list.append((sco_var, i))
self._grb_to_var_ind[grb.VarName] = index_val_list
if DEBUG: self.check_sync()
return sco_var
#@profile
def check_grb_sync(self, grb_name):
for var, i in self._grb_to_var_ind[grb_name]:
print var._grb_vars[i][0].VarName, var._value[i]
#@profile
def check_sync(self):
"""
This function checks whether all sco variable are synchronized
"""
grb_val_map = {}
correctness = True
for grb_name in self._grb_to_var_ind.keys():
for var, i in self._grb_to_var_ind[grb_name]:
try:
correctness = np.allclose(grb_val_map[grb_name], var._value[i], equal_nan=True)
except KeyError:
grb_val_map[grb_name] = var._value[i]
except:
print "something went wrong"
import ipdb; ipdb.set_trace()
if not correctness:
import ipdb; ipdb.set_trace()
def reset_variable(self):
self.grb_init_mapping = {}
self.var_list = []
self._grb_to_var_ind = {}
def monitor_update(self, plan, update_values, callback=None, n_resamples=5, active_ts=None, verbose=False):
print 'Resolving after environment update...\n'
if callback is not None: viewer = callback()
if active_ts==None:
active_ts = (0, plan.horizon-1)
plan.save_free_attrs()
model = grb.Model()
model.params.OutputFlag = 0
self._prob = Prob(model, callback=callback)
# _free_attrs is paied attentioned in here
self._spawn_parameter_to_ll_mapping(model, plan, active_ts)
model.update()
self._bexpr_to_pred = {}
tol = 1e-3
obj_bexprs = []
rs_obj = self._update(plan, update_values)
obj_bexprs.extend(self._get_transfer_obj(plan, self.transfer_norm))
self._add_all_timesteps_of_actions(plan, priority=MAX_PRIORITY,
add_nonlin=True, active_ts= active_ts, verbose=verbose)
obj_bexprs.extend(rs_obj)
self._add_obj_bexprs(obj_bexprs)
initial_trust_region_size = 1e-2
solv = Solver()
solv.initial_trust_region_size = initial_trust_region_size
solv.initial_penalty_coeff = self.init_penalty_coeff
solv.max_merit_coeff_increases = self.max_merit_coeff_increases
success = solv.solve(self._prob, method='penalty_sqp', tol=tol, verbose=verbose)
self._update_ll_params()
if DEBUG: assert not plan.has_nan(active_ts)
plan.restore_free_attrs()
self.reset_variable()
print "monitor_update\n"
return success
def _update(self, plan, update_values):
bexprs = []
for val, attr_inds in update_values:
if val is not None:
for p in attr_inds:
## get the ll_param for p and gurobi variables
ll_p = self._param_to_ll[p]
n_vals, i = 0, 0
grb_vars = []
for attr, ind_arr, t in attr_inds[p]:
for j, grb_var in enumerate(getattr(ll_p, attr)[ind_arr, t-ll_p.active_ts[0]].flatten()):
Q = np.eye(1)
A = -2*val[p][i+j]*np.ones((1, 1))
b = np.ones((1, 1))*np.power(val[p][i+j], 2)
resample_coeff = self.rs_coeff/float(plan.horizon)
# QuadExpr is 0.5*x^Tx + Ax + b
quad_expr = QuadExpr(2*Q*resample_coeff, A*resample_coeff, b*resample_coeff)
v_arr = np.array([grb_var]).reshape((1, 1), order='F')
init_val = np.ones((1, 1))*val[p][i+j]
sco_var = self.create_variable(v_arr, np.array([val[p][i+j]]).reshape((1, 1)), save = True)
bexpr = BoundExpr(quad_expr, sco_var)
bexprs.append(bexpr)
i += len(ind_arr)
return bexprs
#@profile
def _resample(self, plan, preds, sample_all = False):
"""
This function first calls fail predicate's resample function,
then, uses the resampled value to create a square difference cost
function e(x) = |x - rs_val|^2 that will be minimized later.
rs_val is the resampled value
"""
bexprs = []
val, attr_inds = None, None
pred_type = {}
for negated, pred, t in preds:
## returns a vector of new values and an
## attr_inds (OrderedDict) that gives the mapping
## to parameter attributes
if pred_type.get(pred.get_type, False):
continue
val, attr_inds = pred.resample(negated, t, plan)
if val is not None: pred_type[pred.get_type] = True
## if no resample defined for that pred, continue
if val is not None:
for p in attr_inds:
## get the ll_param for p and gurobi variables
ll_p = self._param_to_ll[p]
n_vals, i = 0, 0
grb_vars = []
for attr, ind_arr, t in attr_inds[p]:
for j, grb_var in enumerate(getattr(ll_p, attr)[ind_arr, t-ll_p.active_ts[0]].flatten()):
Q = np.eye(1)
A = -2*val[p][i+j]*np.ones((1, 1))
b = np.ones((1, 1))*np.power(val[p][i+j], 2)
resample_coeff = self.rs_coeff/float(plan.horizon)
# QuadExpr is 0.5*x^Tx + Ax + b
quad_expr = QuadExpr(2*Q*resample_coeff, A*resample_coeff, b*resample_coeff)
v_arr = np.array([grb_var]).reshape((1, 1), order='F')
init_val = np.ones((1, 1))*val[p][i+j]
sco_var = self.create_variable(v_arr, np.array([val[p][i+j]]).reshape((1, 1)), save = True)
bexpr = BoundExpr(quad_expr, sco_var)
bexprs.append(bexpr)
i += len(ind_arr)
if not sample_all:
break
return bexprs
#@profile
def _add_pred_dict(self, pred_dict, effective_timesteps, add_nonlin=True,
priority=MAX_PRIORITY, verbose=False):
"""
This function creates constraints for the predicate and added to
Prob class in sco.
"""
## for debugging
ignore_preds = []
priority = np.maximum(priority, 0)
if not pred_dict['hl_info'] == "hl_state":
start, end = pred_dict['active_timesteps']
active_range = range(start, end+1)
if verbose:
print "pred being added: ", pred_dict
print active_range, effective_timesteps
negated = pred_dict['negated']
pred = pred_dict['pred']
if pred.get_type() in ignore_preds:
return
if pred.priority > priority: return
if DEBUG: assert isinstance(pred, common_predicates.ExprPredicate)
expr = pred.get_expr(negated)
if expr is not None:
if add_nonlin or isinstance(expr.expr, AffExpr):
for t in effective_timesteps:
if t in active_range:
if verbose:
print "expr being added at time ", t
var = self._spawn_sco_var_for_pred(pred, t)
bexpr = BoundExpr(expr, var)
# TODO: REMOVE line below, for tracing back predicate for debugging.
if DEBUG:
bexpr.source = (negated, pred, t)
self._bexpr_to_pred[bexpr] = (negated, pred, t)
groups = ['all']
if self.early_converge:
## this will check for convergence per parameter
## this is good if e.g., a single trajectory quickly
## gets stuck
groups.extend([param.name for param in pred.params])
self._prob.add_cnt_expr(bexpr, groups)
#@profile
def _add_first_and_last_timesteps_of_actions(self, plan, priority = MAX_PRIORITY,
add_nonlin=False, active_ts=None, verbose=False):
"""
Adding only non-linear constraints on the first and last timesteps of each action.
"""
if active_ts is None:
active_ts = (0, plan.horizon-1)
for action in plan.actions:
action_start, action_end = action.active_timesteps
## only add an action
if action_start >= active_ts[1] and action_start > active_ts[0]: continue
if action_end < active_ts[0]: continue
for pred_dict in action.preds:
if action_start >= active_ts[0]:
self._add_pred_dict(pred_dict, [action_start],
priority=priority, add_nonlin=add_nonlin, verbose=verbose)
if action_end <= active_ts[1]:
self._add_pred_dict(pred_dict, [action_end],
priority=priority, add_nonlin=add_nonlin, verbose=verbose)
## add all of the linear ineqs
timesteps = range(max(action_start+1, active_ts[0]),
min(action_end, active_ts[1]))
for pred_dict in action.preds:
self._add_pred_dict(pred_dict, timesteps, add_nonlin=False,
priority=priority, verbose=verbose)
#@profile
def _add_all_timesteps_of_actions(self, plan, priority=MAX_PRIORITY,
add_nonlin=True, active_ts=None, verbose=False):
"""
This function adds both linear and non-linear predicates from
actions that are active within the range of active_ts.
"""
if active_ts==None:
active_ts = (0, plan.horizon-1)
for action in plan.actions:
action_start, action_end = action.active_timesteps
if action_start >= active_ts[1] and action_start > active_ts[0]: continue
if action_end < active_ts[0]: continue
timesteps = range(max(action_start, active_ts[0]),
min(action_end, active_ts[1])+1)
for pred_dict in action.preds:
self._add_pred_dict(pred_dict, timesteps, priority=priority,
add_nonlin=add_nonlin, verbose=verbose)
#@profile
def _update_ll_params(self):
"""
update plan's parameters from low level grb_vars.
expected to be called after each optimization.
"""
for ll_param in self._param_to_ll.values():
ll_param.update_param()
if self.child_solver:
self.child_solver._update_ll_params()
#@profile
def _spawn_parameter_to_ll_mapping(self, model, plan, active_ts=None):
"""
This function creates low level parameters for each parameter in the plan,
initialized he corresponding grb_vars for each attributes in each timestep,
update the grb models
adds in equality constraints,
construct a dictionary as param-to-ll_param mapping.
"""
if active_ts == None:
active_ts=(0, plan.horizon-1)
horizon = active_ts[1] - active_ts[0] + 1
self._param_to_ll = {}
self.ll_start = active_ts[0]
for param in plan.params.values():
ll_param = LLParam(model, param, horizon, active_ts)
ll_param.create_grb_vars()
self._param_to_ll[param] = ll_param
model.update()
for ll_param in self._param_to_ll.values():
ll_param.batch_add_cnts()
#@profile
def _add_obj_bexprs(self, obj_bexprs):
"""
This function adds objective bounded expressions to the Prob class
in sco.
"""
for bexpr in obj_bexprs:
self._prob.add_obj_expr(bexpr)
#@profile
def _get_trajopt_obj(self, plan, active_ts=None):
"""
This function selects parameter of type Robot and Can and returns
the expression e(x) = |Px|^2
Which optimize trajectory so that robot and can's attributes in
current timestep is close to that of next timestep.
forming a straight line between each end points.
Where P is the KT x KT matrix, where Px is the difference of
value in current timestep compare to next timestep
"""
if active_ts == None:
active_ts = (0, plan.horizon-1)
start, end = active_ts
traj_objs = []
for param in plan.params.values():
if param not in self._param_to_ll:
continue
if isinstance(param, Object):
for attr_name in param.__dict__.iterkeys():
attr_type = param.get_attr_type(attr_name)
if issubclass(attr_type, Vector):
T = end - start + 1
K = attr_type.dim
attr_val = getattr(param, attr_name)
KT = K*T
v = -1 * np.ones((KT - K, 1))
d = np.vstack((np.ones((KT - K, 1)), np.zeros((K, 1))))
# [:,0] allows numpy to see v and d as one-dimensional so
# that numpy will create a diagonal matrix with v and d as a diagonal
P = np.diag(v[:, 0], K) + np.diag(d[:, 0])
Q = np.dot(np.transpose(P), P)
Q *= self.trajopt_coeff/float(plan.horizon)
quad_expr = None
quad_expr = QuadExpr(Q, np.zeros((1,KT)), np.zeros((1,1)))
param_ll = self._param_to_ll[param]
ll_attr_val = getattr(param_ll, attr_name)
param_ll_grb_vars = ll_attr_val.reshape((KT, 1), order='F')
attr_val = getattr(param, attr_name)
init_val = attr_val[:, start:end+1].reshape((KT, 1), order='F')
sco_var = self.create_variable(param_ll_grb_vars, init_val)
bexpr = BoundExpr(quad_expr, sco_var)
traj_objs.append(bexpr)
return traj_objs
def _solve_opt_prob(self, plan, priority, callback=None, init=True,
active_ts=None, verbose=False, resample=False, smoothing = False):
if callback is not None: viewer = callback()
self.plan = plan
vehicle = plan.params['vehicle']
if active_ts==None:
active_ts = (0, plan.horizon-1)
plan.save_free_attrs()
model = grb.Model()
model.params.OutputFlag = 0
self._prob = Prob(model, callback=callback)
self._spawn_parameter_to_ll_mapping(model, plan, active_ts)
model.update()
initial_trust_region_size = self.initial_trust_region_size
if resample:
tol = 1e-3
def variable_helper():
error_bin = []
for sco_var in self._prob._vars:
for grb_var, val in zip(sco_var.get_grb_vars(), sco_var.get_value()):
grb_name = grb_var[0].VarName
one, two = grb_name.find('-'), grb_name.find('-(')
param_name = grb_name[1:one]
attr = grb_name[one+1:two]
index = eval(grb_name[two+1:-1])
param = plan.params[param_name]
if not np.allclose(val, getattr(param, attr)[index]):
error_bin.append((grb_name, val, getattr(param, attr)[index]))
if len(error_bin) != 0:
print "something wrong"
if DEBUG: import ipdb; ipdb.set_trace()
"""
When Optimization fails, resample new values for certain timesteps
of the trajectory and solver as initialization
"""
obj_bexprs = []
## this is an objective that places
## a high value on matching the resampled values
failed_preds = plan.get_failed_preds(active_ts = active_ts, priority=priority, tol = tol)
rs_obj = self._resample(plan, failed_preds, sample_all = True)
# import ipdb; ipdb.set_trace()
# _get_transfer_obj returns the expression saying the current trajectory should be close to it's previous trajectory.
# obj_bexprs.extend(self._get_trajopt_obj(plan, active_ts))
obj_bexprs.extend(self._get_transfer_obj(plan, self.transfer_norm))
self._add_all_timesteps_of_actions(plan, priority=priority,
add_nonlin=False, active_ts= active_ts, verbose=verbose)
obj_bexprs.extend(rs_obj)
self._add_obj_bexprs(obj_bexprs)
initial_trust_region_size = 1e3
# import ipdb; ipdb.set_trace()
else:
self._bexpr_to_pred = {}
if priority == -2:
"""
Initialize an linear trajectory while enforceing the linear constraints in the intermediate step.
"""
obj_bexprs = self._get_trajopt_obj(plan, active_ts)
self._add_obj_bexprs(obj_bexprs)
self._add_first_and_last_timesteps_of_actions(plan,
priority=MAX_PRIORITY, active_ts=active_ts, verbose=verbose, add_nonlin=False)
tol = 1e-1
initial_trust_region_size = 1e3
elif priority == -1:
"""
Solve the optimization problem while enforcing every constraints.
"""
obj_bexprs = self._get_trajopt_obj(plan, active_ts)
self._add_obj_bexprs(obj_bexprs)
self._add_first_and_last_timesteps_of_actions(plan,
priority=MAX_PRIORITY, active_ts=active_ts, verbose=verbose,
add_nonlin=True)
tol = 1e-1
elif priority >= 0:
obj_bexprs = self._get_trajopt_obj(plan, active_ts)
self._add_obj_bexprs(obj_bexprs)
self._add_all_timesteps_of_actions(plan, priority=priority, add_nonlin=True,
active_ts=active_ts, verbose=verbose)
tol=1e-3
solv = Solver()
solv.initial_trust_region_size = initial_trust_region_size
if smoothing:
solv.initial_penalty_coeff = self.smooth_penalty_coeff
else:
solv.initial_penalty_coeff = self.init_penalty_coeff
solv.max_merit_coeff_increases = self.max_merit_coeff_increases
success = solv.solve(self._prob, method='penalty_sqp', tol=tol, verbose=verbose)
if priority == MAX_PRIORITY:
success = success or len(plan.get_failed_preds(tol=tol, active_ts=active_ts, priority=priority)) == 0
self._update_ll_params()
if DEBUG: assert not plan.has_nan(active_ts)
if resample:
# During resampling phases, there must be changes added to sampling_trace
if len(plan.sampling_trace) > 0 and 'reward' not in plan.sampling_trace[-1]:
reward = 0
if len(plan.get_failed_preds(active_ts = active_ts, priority=priority)) == 0:
reward = len(plan.actions)
else:
failed_t = plan.get_failed_pred(active_ts=(0,active_ts[1]), priority=priority)[2]
for i in range(len(plan.actions)):
if failed_t > plan.actions[i].active_timesteps[1]:
reward += 1
plan.sampling_trace[-1]['reward'] = reward
##Restore free_attrs values
plan.restore_free_attrs()
self.reset_variable()
print "priority: {}\n".format(priority)
return success
class CanSolver(LLSolver):
def __init__(self, early_converge=False):
self.transfer_coeff = 1e1
self.rs_coeff = 1e6
self.initial_trust_region_size = 1e-1
self.init_penalty_coeff = 1e1
# self.init_penalty_coeff = 1e5
self.max_merit_coeff_increases = 5
self._param_to_ll = {}
self.early_converge = early_converge
self.child_solver = None
def backtrack_solve(self, plan, callback=None, anum=0, verbose=False):
if anum > len(plan.actions) - 1:
return True
a = plan.actions[anum]
active_ts = a.active_timesteps
inits = {}
# Moveto -> Grasp -> Movetoholding -> Putdown
if a.name == "moveto":
## moveto: (?robot - Robot ?start - RobotPose ?end - RobotPose)
## find possible values for the final pose
rs_param = a.params[2]
elif a.name == "movetoholding":
## movetoholding: (?robot - Robot ?start - RobotPose ?end - RobotPose ?c - Can)
## find possible values for the final pose
rs_param = a.params[3]
elif a.name == "grasp":
## grasp: (?robot - Robot ?can - Can ?target - Target ?sp - RobotPose ?ee - EEPose ?ep - RobotPose)
## sample the ee_pose
rs_param = a.params[5]
elif a.name == "putdown":
## putdown: (?robot - Robot ?can - Can ?target - Target ?sp - RobotPose ?ee - EEPose ?ep - RobotPose)
rs_param = a.params[4]
else:
raise NotImplemented
def recursive_solve():
## don't optimize over any params that are already set
old_params_free = {}
for p in plan.params.itervalues():
old_param_map = {}
if p.is_symbol():
if p not in a.params:
continue
for attr in attr_map[getattr(p, "_type")]:
old_param_map[attr] = p._free_attrs[attr].copy()
p._free_attrs[attr][:] = 0
else:
for attr in attr_map[getattr(p, "_type")]:
old_param_map[attr] = p._free_attrs[attr][:, active_ts[1]].copy()
p._free_attrs[attr][:, active_ts[1]] = 0
old_params_free[p] = old_param_map
self.child_solver = CanSolver()
if self.child_solver.backtrack_solve(plan, callback=callback, anum=anum + 1, verbose=verbose):
return True
## reset free_attrs
for p in a.params:
if p.is_symbol():
if p not in a.params:
continue
for attr in attr_map[getattr(p, "_type")]:
p._free_attrs[attr] = old_params_free[p][attr]
else:
for attr in attr_map[getattr(p, "_type")]:
p._free_attrs[attr][:, active_ts[1]] = old_params_free[p][attr]
return False
if rs_param.is_fixed(attr_map[getattr(rs_param, "_type")]):
## this parameter is fixed
if callback is not None:
callback_a = lambda: callback(a)
else:
callback_a = None
self.child_solver = CanSolver()
success = self.child_solver.solve(
plan, callback=callback_a, n_resamples=0, active_ts=active_ts, verbose=verbose, force_init=True
)
if not success:
## if planning fails we're done
return False
## no other options, so just return here
return recursive_solve()
## so that this won't be optimized over
rs_free = {}
for attr in attr_map[getattr(rs_param, "_type")]:
rs_free[attr] = rs_param._free_attrs[attr].copy()
rs_param._free_attrs[attr][:] = 0
# Target is needed to sample the ee_pose, ee_pose is needed to sample the robot_pose
# InContact, Robot, EEPose, Target
targets = plan.get_param("InContact", 2, {1: rs_param}, negated=False)
# EEReachable, Robot, RobotPose, EEPose
robot_pose = plan.get_param("EEReachable", 1, {2: rs_param}, negated=False)
if len(targets) > 1:
import pdb
pdb.set_trace()
if callback is not None:
callback_a = lambda: callback(a)
else:
callback_a = None
robot_poses = []
if len(targets) == 0 or np.all([targets[0]._free_attrs[attr] for attr in attr_map[targets[0]]]):
## sample 4 possible poses
coords = list(itertools.product(range(WIDTH), range(HEIGHT)))
random.shuffle(coords)
robot_poses = [np.array(x)[:, None] for x in coords[:4]]
elif np.any([targets[0]._free_attrs[attr] for attr in attr_map[targets[0]]]):
## there shouldn't be only some free_attrs set
raise NotImplementedError
else:
grasp_dirs = [np.array([0, -1]), np.array([1, 0]), np.array([0, 1]), np.array([-1, 0])]
grasp_len = plan.params["pr2"].geom.radius + targets[0].geom.radius
for g_dir in grasp_dirs:
grasp = (g_dir * grasp_len).reshape((2, 1))
robot_poses.append(targets[0].value + grasp)
for rp in robot_poses:
rs_param.value = rp
success = False
self.child_solver = CanSolver()
success = self.child_solver.solve(
plan, callback=callback_a, n_resamples=0, active_ts=active_ts, verbose=verbose, force_init=True
)
if success:
if recursive_solve():
break
else:
success = False
for attr in attr_map[getattr(rs_param, "_type")]:
rs_param._free_attrs[attr] = rs_free[attr]
return success
def sample_ee_from_target(self, target):
"""
Sample all possible EE Pose that pr2 can grasp with
target: parameter of type Target
return: list of tuple in the format of (ee_pos, ee_rot)
"""
possible_ee_poses = []
targ_pos, targ_rot = target.value.flatten(), target.rotation.flatten()
ee_pos = targ_pos.copy()
target_trans = OpenRAVEBody.transform_from_obj_pose(targ_pos, targ_rot)
# rotate can's local z-axis by the amount of linear spacing between 0 to 2pi
angle_range = np.linspace(0, np.pi * 2, num=SAMPLE_SIZE)
for rot in angle_range:
target_trans = OpenRAVEBody.transform_from_obj_pose(targ_pos, targ_rot)
# rotate new ee_pose around can's rotation axis
rot_mat = matrixFromAxisAngle([0, 0, rot])
ee_trans = target_trans.dot(rot_mat)
ee_rot = OpenRAVEBody.obj_pose_from_transform(ee_trans)[3:]
possible_ee_poses.append((ee_pos, ee_rot))
return possible_ee_poses
def sample_rp_from_ee(self, env, rs_param, ee_pose):
"""
Using Inverse Kinematics to solve all possible robot armpose and basepose that reaches this ee_pose
robot_pose: list of full robot pose with format of(basepose, backHeight, lArmPose, lGripper, rArmPose, rGripper)
"""
ee_pos, ee_rot = ee_pose[0], ee_pose[1]
base_raidus = np.linspace(-1, 1, num=BASE_SAMPLE_SIZE)
poses = list(itertools.product(base_raidus, base_raidus))
# sample
dummy_body = OpenRAVEBody(env, rs_param.name, PR2())
possible_robot_pose = []
for pos in poses:
bp = pos + ee_pos[:2]
vec = ee_pos[:2] - bp
vec = vec / np.linalg.norm(vec)
theta = math.atan2(vec[1], vec[0])
bp = np.array([bp[0], bp[1], theta])
dummy_body.set_pose(bp)
arm_poses = robot_body.ik_arm_pose(ee_pos, ee_rot)
for arm_pose in arm_poses:
possible_robot_pose.append(
(bp, ee_arm_poses[0], rs_param.lArmPose, rs_param.lGripper, ee_arm_poses[1:], rs_param.rGripper)
)
import ipdb
ipdb.set_trace()
dummy_body.delete()
return possible_robot_pose
def closest_arm_pose(self, arm_poses, cur_arm_poses):
pass
def solve(self, plan, callback=None, n_resamples=5, active_ts=None, verbose=False, force_init=False):
success = False
if force_init or not plan.initialized:
## solve at priority -1 to get an initial value for the parameters
self._solve_opt_prob(plan, priority=-1, callback=callback, active_ts=active_ts, verbose=verbose)
plan.initialized = True
success = self._solve_opt_prob(plan, priority=1, callback=callback, active_ts=active_ts, verbose=verbose)
if success:
return success
for _ in range(n_resamples):
## refinement loop
## priority 0 resamples the first failed predicate in the plan
## and then solves a transfer optimization that only includes linear constraints
self._solve_opt_prob(plan, priority=0, callback=callback, active_ts=active_ts, verbose=verbose)
success = self._solve_opt_prob(plan, priority=1, callback=callback, active_ts=active_ts, verbose=verbose)
if success:
return success
return success
def _solve_opt_prob(
self, plan, priority, callback=None, init=True, init_from_prev=False, active_ts=None, verbose=False
):
## active_ts is the inclusive timesteps to include
## in the optimization
if active_ts == None:
active_ts = (0, plan.horizon - 1)
model = grb.Model()
model.params.OutputFlag = 0
self._prob = Prob(model, callback=callback)
# param_to_ll_old = self._param_to_ll.copy()
self._spawn_parameter_to_ll_mapping(model, plan, active_ts)
model.update()
self._bexpr_to_pred = {}
if priority == -1:
obj_bexprs = self._get_trajopt_obj(plan, active_ts)
self._add_obj_bexprs(obj_bexprs)
self._add_first_and_last_timesteps_of_actions(plan, priority=-1, active_ts=active_ts, verbose=verbose)
tol = 1e-1
elif priority == 0:
## this should only get called with a full plan for now
assert active_ts == (0, plan.horizon - 1)
failed_preds = plan.get_failed_preds()
## this is an objective that places
## a high value on matching the resampled values
obj_bexprs = self._resample(plan, failed_preds)
## solve an optimization movement primitive to
## transfer current trajectories
obj_bexprs.extend(self._get_transfer_obj(plan, "min-vel"))
self._add_obj_bexprs(obj_bexprs)
# self._add_first_and_last_timesteps_of_actions(
# plan, priority=0, add_nonlin=False)
self._add_first_and_last_timesteps_of_actions(plan, priority=-1, add_nonlin=True, verbose=verbose)
self._add_all_timesteps_of_actions(plan, priority=0, add_nonlin=False, verbose=verbose)
tol = 1e-1
elif priority == 1:
obj_bexprs = self._get_trajopt_obj(plan, active_ts)
self._add_obj_bexprs(obj_bexprs)
self._add_all_timesteps_of_actions(plan, priority=1, add_nonlin=True, active_ts=active_ts, verbose=verbose)
tol = 1e-4
solv = Solver()
solv.initial_trust_region_size = self.initial_trust_region_size
solv.initial_penalty_coeff = self.init_penalty_coeff
solv.max_merit_coeff_increases = self.max_merit_coeff_increases
success = solv.solve(self._prob, method="penalty_sqp", tol=tol, verbose=True)
import ipdb
ipdb.set_trace()
self._update_ll_params()
return success
def _get_transfer_obj(self, plan, norm):
transfer_objs = []
if norm == "min-vel":
for param in plan.params.values():
if param._type in ["Robot", "Can"]:
T = plan.horizon
K = 2
pose = param.pose
assert (K, T) == pose.shape
KT = K * T
v = -1 * np.ones((KT - K, 1))
d = np.vstack((np.ones((KT - K, 1)), np.zeros((K, 1))))
# [:,0] allows numpy to see v and d as one-dimensional so
# that numpy will create a diagonal matrix with v and d as a diagonal
P = np.diag(v[:, 0], K) + np.diag(d[:, 0])
Q = np.dot(np.transpose(P), P)
cur_pose = pose.reshape((KT, 1), order="F")
A = -2 * cur_pose.T.dot(Q)
b = cur_pose.T.dot(Q.dot(cur_pose))
# QuadExpr is 0.5*x^Tx + Ax + b
quad_expr = QuadExpr(2 * self.transfer_coeff * Q, self.transfer_coeff * A, self.transfer_coeff * b)
ll_param = self._param_to_ll[param]
ll_grb_vars = ll_param.pose.reshape((KT, 1), order="F")
bexpr = BoundExpr(quad_expr, Variable(ll_grb_vars, cur_pose))
transfer_objs.append(bexpr)
else:
raise NotImplemented
return transfer_objs
def _resample(self, plan, preds):
val, attr_inds = None, None
for negated, pred, t in preds:
## returns a vector of new values and an
## attr_inds (OrderedDict) that gives the mapping
## to parameter attributes
val, attr_inds = pred.resample(negated, t, plan)
## no resample defined for that pred
if val is not None:
break
if val is None:
return None
t_local = t
bexprs = []
i = 0
for p in attr_inds:
## get the ll_param for p and gurobi variables
ll_p = self._param_to_ll[p]
if p.is_symbol():
t_local = 0
n_vals = 0
grb_vars = []
for attr, ind_arr in attr_inds[p]:
n_vals += len(ind_arr)
grb_vars.extend(list(getattr(ll_p, attr)[ind_arr, t_local]))
for j, grb_var in enumerate(grb_vars):
## create an objective saying stay close to this value
## e(x) = x^2 - 2*val[i+j]*x + val[i+j]^2
Q = np.eye(1)
A = -2 * val[i + j] * np.ones((1, 1))
b = np.ones((1, 1)) * np.power(val[i + j], 2)
# QuadExpr is 0.5*x^Tx + Ax + b
quad_expr = QuadExpr(2 * Q * self.rs_coeff, A * self.rs_coeff, b * self.rs_coeff)
v_arr = np.array([grb_var]).reshape((1, 1), order="F")
init_val = np.ones((1, 1)) * val[i + j]
bexpr = BoundExpr(quad_expr, Variable(v_arr, val[i + j].reshape((1, 1))))
bexprs.append(bexpr)
i += n_vals
return bexprs
def _add_pred_dict(self, pred_dict, effective_timesteps, add_nonlin=True, priority=MAX_PRIORITY, verbose=False):
## for debugging
ignore_preds = []
priority = np.maximum(priority, 0)
if not pred_dict["hl_info"] == "hl_state":
start, end = pred_dict["active_timesteps"]
active_range = range(start, end + 1)
if verbose:
print "pred being added: ", pred_dict
print active_range, effective_timesteps
negated = pred_dict["negated"]
pred = pred_dict["pred"]
if pred.get_type() in ignore_preds:
return
if pred.priority > priority:
return
assert isinstance(pred, common_predicates.ExprPredicate)
expr = pred.get_expr(negated)
for t in effective_timesteps:
if t in active_range:
if expr is not None:
if add_nonlin or isinstance(expr.expr, AffExpr):
if verbose:
print "expr being added at time ", t
var = self._spawn_sco_var_for_pred(pred, t)
bexpr = BoundExpr(expr, var)
# TODO: REMOVE line below, for tracing back predicate for debugging.
bexpr.pred = pred
self._bexpr_to_pred[bexpr] = (negated, pred, t)
groups = ["all"]
if self.early_converge:
## this will check for convergence per parameter
## this is good if e.g., a single trajectory quickly
## gets stuck
groups.extend([param.name for param in pred.params])
self._prob.add_cnt_expr(bexpr, groups)
def _add_first_and_last_timesteps_of_actions(
self, plan, priority=MAX_PRIORITY, add_nonlin=False, active_ts=None, verbose=False
):
if active_ts == None:
active_ts = (0, plan.horizon - 1)
for action in plan.actions:
action_start, action_end = action.active_timesteps
## only add an action
if action_start >= active_ts[1]:
continue
if action_end < active_ts[0]:
continue
for pred_dict in action.preds:
if action_start >= active_ts[0]:
self._add_pred_dict(
pred_dict, [action_start], priority=priority, add_nonlin=add_nonlin, verbose=verbose
)
if action_end <= active_ts[1]:
self._add_pred_dict(
pred_dict, [action_end], priority=priority, add_nonlin=add_nonlin, verbose=verbose
)
## add all of the linear ineqs
timesteps = range(max(action_start + 1, active_ts[0]), min(action_end, active_ts[1]))
for pred_dict in action.preds:
self._add_pred_dict(pred_dict, timesteps, add_nonlin=False, priority=priority, verbose=verbose)
def _add_all_timesteps_of_actions(
self, plan, priority=MAX_PRIORITY, add_nonlin=True, active_ts=None, verbose=False
):
if active_ts == None:
active_ts = (0, plan.horizon - 1)
for action in plan.actions:
action_start, action_end = action.active_timesteps
if action_start > active_ts[1]:
continue
if action_end < active_ts[0]:
continue
timesteps = range(max(action_start, active_ts[0]), min(action_end, active_ts[1]) + 1)
for pred_dict in action.preds:
self._add_pred_dict(pred_dict, timesteps, priority=priority, add_nonlin=add_nonlin, verbose=verbose)
def _update_ll_params(self):
for ll_param in self._param_to_ll.values():
ll_param.update_param()
if self.child_solver:
self.child_solver._update_ll_params()
def _spawn_parameter_to_ll_mapping(self, model, plan, active_ts=None):
if active_ts == None:
active_ts = (0, plan.horizon - 1)
horizon = active_ts[1] - active_ts[0] + 1
self._param_to_ll = {}
self.ll_start = active_ts[0]
for param in plan.params.values():
ll_param = LLParam(model, param, horizon, active_ts)
ll_param.create_grb_vars()
self._param_to_ll[param] = ll_param
model.update()
for ll_param in self._param_to_ll.values():
ll_param.batch_add_cnts()
def _add_obj_bexprs(self, obj_bexprs):
for bexpr in obj_bexprs:
self._prob.add_obj_expr(bexpr)
def _get_trajopt_obj(self, plan, active_ts=None):
if active_ts == None:
active_ts = (0, plan.horizon - 1)
start, end = active_ts
traj_objs = []
for param in plan.params.values():
if param not in self._param_to_ll:
continue
if param._type in ["Robot", "Can"]:
for attr_name in param.__dict__.iterkeys():
attr_type = param.get_attr_type(attr_name)
if issubclass(attr_type, Vector):
T = end - start + 1
K = attr_type.dim
attr_val = getattr(param, attr_name)
KT = K * T
v = -1 * np.ones((KT - K, 1))
d = np.vstack((np.ones((KT - K, 1)), np.zeros((K, 1))))
# [:,0] allows numpy to see v and d as one-dimensional so
# that numpy will create a diagonal matrix with v and d as a diagonal
P = np.diag(v[:, 0], K) + np.diag(d[:, 0])
Q = np.dot(np.transpose(P), P)
quad_expr = QuadExpr(Q, np.zeros((1, KT)), np.zeros((1, 1)))
robot_ll = self._param_to_ll[param]
ll_attr_val = getattr(robot_ll, attr_name)
robot_ll_grb_vars = ll_attr_val.reshape((KT, 1), order="F")
# init_val = attr_val[:, start:end+1].reshape((KT, 1), order='F')
# bexpr = BoundExpr(quad_expr, Variable(robot_ll_grb_vars, init_val))
bexpr = BoundExpr(quad_expr, Variable(robot_ll_grb_vars))
traj_objs.append(bexpr)
return traj_objs
def _spawn_sco_var_for_pred(self, pred, t):
i = 0
x = np.empty(pred.x_dim, dtype=object)
v = np.empty(pred.x_dim)
for p in pred.attr_inds:
for attr, ind_arr in pred.attr_inds[p]:
n_vals = len(ind_arr)
ll_p = self._param_to_ll[p]
if p.is_symbol():
x[i : i + n_vals] = getattr(ll_p, attr)[ind_arr, 0]
v[i : i + n_vals] = getattr(p, attr)[ind_arr, 0]
else:
x[i : i + n_vals] = getattr(ll_p, attr)[ind_arr, t - self.ll_start]
v[i : i + n_vals] = getattr(p, attr)[ind_arr, t]
i += n_vals
if pred.dynamic:
## include the variables from the next time step
for p in pred.attr_inds:
for attr, ind_arr in pred.attr_inds[p]:
n_vals = len(ind_arr)
ll_p = self._param_to_ll[p]
if p.is_symbol():
x[i : i + n_vals] = getattr(ll_p, attr)[ind_arr, 0]
v[i : i + n_vals] = getattr(p, attr)[ind_arr, 0]
else:
x[i : i + n_vals] = getattr(ll_p, attr)[ind_arr, t + 1 - self.ll_start]
v[i : i + n_vals] = getattr(p, attr)[ind_arr, t]
i += n_vals
assert i >= pred.x_dim
x = x.reshape((pred.x_dim, 1))
v = v.reshape((pred.x_dim, 1))
return Variable(x, v)
def _solve_opt_prob(self, plan, priority, callback=None, init=True, active_ts=None,
verbose=False):
## active_ts is the inclusive timesteps to include
## in the optimization
if active_ts==None:
active_ts = (0, plan.horizon-1)
model = grb.Model()
model.params.OutputFlag = 0
self._prob = Prob(model, callback=callback)
# param_to_ll_old = self._param_to_ll.copy()
self._spawn_parameter_to_ll_mapping(model, plan, active_ts)
model.update()
self._bexpr_to_pred = {}
if priority == -2:
obj_bexprs = self._get_trajopt_obj(plan, active_ts)
self._add_obj_bexprs(obj_bexprs)
self._add_first_and_last_timesteps_of_actions(plan, priority=MAX_PRIORITY, active_ts=active_ts, verbose=verbose, add_nonlin=False)
# self._add_all_timesteps_of_actions(plan, priority=1, add_nonlin=False, verbose=verbose)
tol = 1e-1
elif priority == -1:
obj_bexprs = self._get_trajopt_obj(plan, active_ts)
self._add_obj_bexprs(obj_bexprs)
self._add_first_and_last_timesteps_of_actions(plan, priority=MAX_PRIORITY, active_ts=active_ts, verbose=verbose, add_nonlin=True)
tol = 1e-1
elif priority == 0:
## this should only get called with a full plan for now
assert active_ts == (0, plan.horizon-1)
failed_preds = plan.get_failed_preds()
## this is an objective that places
## a high value on matching the resampled values
obj_bexprs = []
obj_bexprs.extend(self._resample(plan, failed_preds))
## solve an optimization movement primitive to
## transfer current trajectories
obj_bexprs.extend(self._get_trajopt_obj(plan, active_ts))
# obj_bexprs.extend(self._get_transfer_obj(plan, 'min-vel'))
self._add_obj_bexprs(obj_bexprs)
# self._add_first_and_last_timesteps_of_actions(
# plan, priority=0, add_nonlin=False)
# self._add_first_and_last_timesteps_of_actions(
# plan, priority=-1, add_nonlin=True, verbose=verbose)
# self._add_all_timesteps_of_actions(
# plan, priority=0, add_nonlin=False, verbose=verbose)
self._add_all_timesteps_of_actions(
plan, priority=1, add_nonlin=True, verbose=verbose)
# self._add_first_and_last_timesteps_of_actions(
# plan, priority=1, add_nonlin=True, verbose=verbose)
# self._add_all_timesteps_of_actions(
# plan, priority=0, add_nonlin=False, verbose=verbose)
tol = 1e-1
elif priority >= 1:
obj_bexprs = self._get_trajopt_obj(plan, active_ts)
self._add_obj_bexprs(obj_bexprs)
self._add_all_timesteps_of_actions(plan, priority=priority, add_nonlin=True, active_ts=active_ts, verbose=verbose)
tol=1e-3
solv = Solver()
solv.initial_trust_region_size = self.initial_trust_region_size
solv.initial_penalty_coeff = self.init_penalty_coeff
solv.max_merit_coeff_increases = self.max_merit_coeff_increases
success = solv.solve(self._prob, method='penalty_sqp', tol=tol, verbose=True)
self._update_ll_params()
print "priority: {}".format(priority)
# if callback is not None: callback(True)
return success
class NAMOSolver(LLSolver):
def __init__(self, early_converge=True, transfer_norm='min-vel'):
self.transfer_coeff = 1e1
self.rs_coeff = 1e6
self.init_penalty_coeff = 1e2
self.child_solver = None
self._param_to_ll = {}
self._failed_groups = []
self.early_converge = early_converge
self.transfer_norm = transfer_norm
def backtrack_solve(self, plan, callback=None, verbose=False):
plan.save_free_attrs()
success = self._backtrack_solve(plan,
callback,
anum=0,
verbose=verbose)
plan.restore_free_attrs()
return success
def _backtrack_solve(self, plan, callback=None, anum=0, verbose=False):
if anum > len(plan.actions) - 1:
return True
a = plan.actions[anum]
active_ts = a.active_timesteps
inits = {}
if a.name == 'moveto':
## find possible values for the final pose
rs_param = a.params[2]
elif a.name == 'movetoholding':
## find possible values for the final pose
rs_param = a.params[2]
elif a.name == 'grasp':
## sample the grasp/grasp_pose
rs_param = a.params[4]
elif a.name == 'putdown':
## sample the end pose
rs_param = a.params[4]
else:
raise NotImplemented
def recursive_solve():
## don't optimize over any params that are already set
old_params_free = {}
for p in plan.params.itervalues():
if p.is_symbol():
if p not in a.params: continue
old_params_free[p] = p._free_attrs['value'].copy()
p._free_attrs['value'][:] = 0
else:
old_params_free[p] = p._free_attrs[
'pose'][:, active_ts[1]].copy()
p._free_attrs['pose'][:, active_ts[1]] = 0
self.child_solver = NAMOSolver()
if self.child_solver._backtrack_solve(plan,
callback=callback,
anum=anum + 1,
verbose=verbose):
return True
## reset free_attrs
for p in a.params:
if p.is_symbol():
if p not in a.params: continue
p._free_attrs['value'] = old_params_free[p]
else:
p._free_attrs['pose'][:, active_ts[1]] = old_params_free[p]
return False
if not np.all(rs_param._free_attrs['value']):
## this parameter is fixed
if callback is not None:
callback_a = lambda: callback(a)
else:
callback_a = None
self.child_solver = NAMOSolver()
success = self.child_solver.solve(plan,
callback=callback_a,
n_resamples=0,
active_ts=active_ts,
verbose=verbose,
force_init=True)
if not success:
## if planning fails we're done
return False
## no other options, so just return here
return recursive_solve()
## so that this won't be optimized over
rs_free = rs_param._free_attrs['value'].copy()
rs_param._free_attrs['value'][:] = 0
targets = plan.get_param('InContact', 2, {1: rs_param}, negated=False)
if len(targets) > 1:
import pdb
pdb.set_trace()
if callback is not None:
callback_a = lambda: callback(a)
else:
callback_a = None
robot_poses = []
if len(targets) == 0 or np.all(targets[0]._free_attrs['value']):
## sample 4 possible poses
coords = list(itertools.product(range(WIDTH), range(HEIGHT)))
random.shuffle(coords)
robot_poses = [np.array(x)[:, None] for x in coords[:4]]
elif np.any(targets[0]._free_attrs['value']):
## there shouldn't be only some free_attrs set
raise NotImplementedError
else:
grasp_dirs = [
np.array([0, -1]),
np.array([1, 0]),
np.array([0, 1]),
np.array([-1, 0])
]
grasp_len = plan.params['pr2'].geom.radius + targets[
0].geom.radius - dsafe
for g_dir in grasp_dirs:
grasp = (g_dir * grasp_len).reshape((2, 1))
robot_poses.append(targets[0].value + grasp)
for rp in robot_poses:
rs_param.value = rp
success = False
self.child_solver = NAMOSolver()
success = self.child_solver.solve(plan,
callback=callback_a,
n_resamples=0,
active_ts=active_ts,
verbose=verbose,
force_init=True)
if success:
if recursive_solve():
break
else:
success = False
rs_param._free_attrs['value'] = rs_free
return success
def solve(self,
plan,
callback=None,
n_resamples=5,
active_ts=None,
verbose=False,
force_init=False):
success = False
if force_init or not plan.initialized:
## solve at priority -1 to get an initial value for the parameters
self._solve_opt_prob(plan,
priority=-1,
callback=callback,
active_ts=active_ts,
verbose=verbose)
plan.initialized = True
success = self._solve_opt_prob(plan,
priority=1,
callback=callback,
active_ts=active_ts,
verbose=verbose)
success = plan.satisfied(active_ts)
if success:
return success
for _ in range(n_resamples):
## refinement loop
## priority 0 resamples the first failed predicate in the plan
## and then solves a transfer optimization that only includes linear constraints
self._solve_opt_prob(plan,
priority=0,
callback=callback,
active_ts=active_ts,
verbose=verbose)
success = self._solve_opt_prob(plan,
priority=1,
callback=callback,
active_ts=active_ts,
verbose=verbose)
success = plan.satisfied(active_ts)
if success:
return success
return success
# @profile
def _solve_opt_prob(self,
plan,
priority,
callback=None,
active_ts=None,
verbose=False,
resample=True):
## active_ts is the inclusive timesteps to include
## in the optimization
if active_ts == None:
active_ts = (0, plan.horizon - 1)
model = grb.Model()
model.params.OutputFlag = 0
self._prob = Prob(model, callback=callback)
# param_to_ll_old = self._param_to_ll.copy()
self._spawn_parameter_to_ll_mapping(model, plan, active_ts)
model.update()
self._bexpr_to_pred = {}
if priority == -1:
obj_bexprs = self._get_trajopt_obj(plan, active_ts)
self._add_obj_bexprs(obj_bexprs)
self._add_first_and_last_timesteps_of_actions(plan,
priority=-1,
active_ts=active_ts,
verbose=verbose)
tol = 1e-2
elif priority == 0:
## this should only get called with a full plan for now
assert active_ts == (0, plan.horizon - 1)
obj_bexprs = []
if resample:
failed_preds = plan.get_failed_preds()
## this is an objective that places
## a high value on matching the resampled values
obj_bexprs.extend(self._resample(plan, failed_preds))
## solve an optimization movement primitive to
## transfer current trajectories
obj_bexprs.extend(self._get_transfer_obj(plan, self.transfer_norm))
self._add_obj_bexprs(obj_bexprs)
# self._add_first_and_last_timesteps_of_actions(
# plan, priority=0, add_nonlin=False)
self._add_first_and_last_timesteps_of_actions(plan,
priority=-1,
add_nonlin=True,
verbose=verbose)
self._add_all_timesteps_of_actions(plan,
priority=0,
add_nonlin=False,
verbose=verbose)
tol = 1e-2
elif priority == 1:
obj_bexprs = self._get_trajopt_obj(plan, active_ts)
self._add_obj_bexprs(obj_bexprs)
self._add_all_timesteps_of_actions(plan,
priority=1,
add_nonlin=True,
active_ts=active_ts,
verbose=verbose)
tol = 1e-4
solv = Solver()
solv.initial_penalty_coeff = self.init_penalty_coeff
success = solv.solve(self._prob,
method='penalty_sqp',
tol=tol,
verbose=verbose)
self._update_ll_params()
self._failed_groups = self._prob.nonconverged_groups
return success
def get_value(self, plan, penalty_coeff=1e0):
model = grb.Model()
model.params.OutputFlag = 0
self._prob = Prob(model)
# param_to_ll_old = self._param_to_ll.copy()
self._spawn_parameter_to_ll_mapping(model, plan)
model.update()
self._bexpr_to_pred = {}
obj_bexprs = self._get_trajopt_obj(plan)
self._add_obj_bexprs(obj_bexprs)
self._add_all_timesteps_of_actions(plan,
priority=1,
add_nonlin=True,
verbose=False)
return self._prob.get_value(penalty_coeff)
def _get_transfer_obj(self, plan, norm):
transfer_objs = []
if norm in ['min-vel', 'l2']:
for param in plan.params.values():
# if param._type in ['Robot', 'Can']:
K = 2
if param.is_symbol():
T = 1
pose = param.value
else:
T = plan.horizon
pose = param.pose
assert (K, T) == pose.shape
KT = K * T
if norm == 'min-vel' and not param.is_symbol():
v = -1 * np.ones((KT - K, 1))
d = np.vstack((np.ones((KT - K, 1)), np.zeros((K, 1))))
# [:,0] allows numpy to see v and d as one-dimensional so
# that numpy will create a diagonal matrix with v and d as a diagonal
P = np.diag(v[:, 0], K) + np.diag(d[:, 0])
Q = np.dot(np.transpose(P), P)
else: ## l2
Q = np.eye(KT)
cur_pose = pose.reshape((KT, 1), order='F')
A = -2 * cur_pose.T.dot(Q)
b = cur_pose.T.dot(Q.dot(cur_pose))
# QuadExpr is 0.5*x^Tx + Ax + b
quad_expr = QuadExpr(2 * self.transfer_coeff * Q,
self.transfer_coeff * A,
self.transfer_coeff * b)
ll_param = self._param_to_ll[param]
if param.is_symbol():
ll_grb_vars = ll_param.value.reshape((KT, 1), order='F')
else:
ll_grb_vars = ll_param.pose.reshape((KT, 1), order='F')
bexpr = BoundExpr(quad_expr, Variable(ll_grb_vars, cur_pose))
transfer_objs.append(bexpr)
elif norm == 'straightline':
return self._get_trajopt_obj(plan)
else:
raise NotImplementedError
return transfer_objs
def _resample(self, plan, preds):
val, attr_inds = None, None
for negated, pred, t in preds:
## returns a vector of new values and an
## attr_inds (OrderedDict) that gives the mapping
## to parameter attributes
if (self.early_converge and self._failed_groups != ['all']
and self._failed_groups != [] and not np.any([
param.name in self._failed_groups
for param in pred.params
])):
## using early converge and one group didn't converge
## but no parameter from this pred in that group
continue
val, attr_inds = pred.resample(negated, t, plan)
## no resample defined for that pred
if val is not None: break
if val is None:
# import pdb; pdb.set_trace()
return []
t_local = t
bexprs = []
i = 0
for p in attr_inds:
## get the ll_param for p and gurobi variables
ll_p = self._param_to_ll[p]
if p.is_symbol(): t_local = 0
n_vals = 0
grb_vars = []
for attr, ind_arr in attr_inds[p]:
n_vals += len(ind_arr)
grb_vars.extend(list(getattr(ll_p, attr)[ind_arr, t_local]))
for j, grb_var in enumerate(grb_vars):
## create an objective saying stay close to this value
## e(x) = x^2 - 2*val[i+j]*x + val[i+j]^2
Q = np.eye(1)
A = -2 * val[i + j] * np.ones((1, 1))
b = np.ones((1, 1)) * np.power(val[i + j], 2)
# QuadExpr is 0.5*x^Tx + Ax + b
quad_expr = QuadExpr(2 * Q * self.rs_coeff, A * self.rs_coeff,
b * self.rs_coeff)
v_arr = np.array([grb_var]).reshape((1, 1), order='F')
init_val = np.ones((1, 1)) * val[i + j]
bexpr = BoundExpr(quad_expr,
Variable(v_arr, val[i + j].reshape((1, 1))))
bexprs.append(bexpr)
i += n_vals
return bexprs
def _add_pred_dict(self,
pred_dict,
effective_timesteps,
add_nonlin=True,
priority=MAX_PRIORITY,
verbose=False):
# verbose=True
## for debugging
ignore_preds = []
priority = np.maximum(priority, 0)
if not pred_dict['hl_info'] == "hl_state":
start, end = pred_dict['active_timesteps']
active_range = range(start, end + 1)
if verbose:
print "pred being added: ", pred_dict
print active_range, effective_timesteps
negated = pred_dict['negated']
pred = pred_dict['pred']
if pred.get_type() in ignore_preds:
return
if pred.priority > priority: return
assert isinstance(pred, common_predicates.ExprPredicate)
expr = pred.get_expr(negated)
for t in effective_timesteps:
if t in active_range:
if expr is not None:
if add_nonlin or isinstance(expr.expr, AffExpr):
if verbose:
print "expr being added at time ", t
var = self._spawn_sco_var_for_pred(pred, t)
bexpr = BoundExpr(expr, var)
self._bexpr_to_pred[bexpr] = (negated, pred, t)
groups = ['all']
if self.early_converge:
## this will check for convergence per parameter
## this is good if e.g., a single trajectory quickly
## gets stuck
groups.extend(
[param.name for param in pred.params])
self._prob.add_cnt_expr(bexpr, groups)
def _add_first_and_last_timesteps_of_actions(self,
plan,
priority=MAX_PRIORITY,
add_nonlin=False,
active_ts=None,
verbose=False):
if active_ts == None:
active_ts = (0, plan.horizon - 1)
for action in plan.actions:
action_start, action_end = action.active_timesteps
## only add an action
if action_start >= active_ts[1] and action_start > active_ts[0]:
continue
if action_end < active_ts[0]: continue
for pred_dict in action.preds:
if action_start >= active_ts[0]:
self._add_pred_dict(pred_dict, [action_start],
priority=priority,
add_nonlin=add_nonlin,
verbose=verbose)
if action_end <= active_ts[1]:
self._add_pred_dict(pred_dict, [action_end],
priority=priority,
add_nonlin=add_nonlin,
verbose=verbose)
## add all of the linear ineqs
timesteps = range(max(action_start + 1, active_ts[0]),
min(action_end, active_ts[1]))
for pred_dict in action.preds:
self._add_pred_dict(pred_dict,
timesteps,
add_nonlin=False,
priority=priority,
verbose=verbose)
def _add_all_timesteps_of_actions(self,
plan,
priority=MAX_PRIORITY,
add_nonlin=True,
active_ts=None,
verbose=False):
if active_ts == None:
active_ts = (0, plan.horizon - 1)
for action in plan.actions:
action_start, action_end = action.active_timesteps
if action_start >= active_ts[1] and action_start > active_ts[0]:
continue
if action_end < active_ts[0]: continue
timesteps = range(max(action_start, active_ts[0]),
min(action_end, active_ts[1]) + 1)
for pred_dict in action.preds:
self._add_pred_dict(pred_dict,
timesteps,
priority=priority,
add_nonlin=add_nonlin,
verbose=verbose)
def _update_ll_params(self):
for ll_param in self._param_to_ll.values():
ll_param.update_param()
if self.child_solver:
self.child_solver._update_ll_params()
def _spawn_parameter_to_ll_mapping(self, model, plan, active_ts=None):
if active_ts == None:
active_ts = (0, plan.horizon - 1)
horizon = active_ts[1] - active_ts[0] + 1
self._param_to_ll = {}
self.ll_start = active_ts[0]
for param in plan.params.values():
ll_param = LLParam(model, param, horizon, active_ts)
ll_param.create_grb_vars()
self._param_to_ll[param] = ll_param
model.update()
for ll_param in self._param_to_ll.values():
ll_param.batch_add_cnts()
def _add_obj_bexprs(self, obj_bexprs):
for bexpr in obj_bexprs:
self._prob.add_obj_expr(bexpr)
def _get_trajopt_obj(self, plan, active_ts=None):
if active_ts == None:
active_ts = (0, plan.horizon - 1)
start, end = active_ts
traj_objs = []
for param in plan.params.values():
if param not in self._param_to_ll:
continue
if param._type in ['Robot', 'Can']:
T = end - start + 1
K = 2
pose = param.pose
KT = K * T
v = -1 * np.ones((KT - K, 1))
d = np.vstack((np.ones((KT - K, 1)), np.zeros((K, 1))))
# [:,0] allows numpy to see v and d as one-dimensional so
# that numpy will create a diagonal matrix with v and d as a diagonal
P = np.diag(v[:, 0], K) + np.diag(d[:, 0])
Q = np.dot(np.transpose(P), P)
Q *= TRAJOPT_COEFF
quad_expr = QuadExpr(Q, np.zeros((1, KT)), np.zeros((1, 1)))
robot_ll = self._param_to_ll[param]
robot_ll_grb_vars = robot_ll.pose.reshape((KT, 1), order='F')
init_val = param.pose[:, start:end + 1].reshape((KT, 1),
order='F')
bexpr = BoundExpr(quad_expr,
Variable(robot_ll_grb_vars, init_val))
traj_objs.append(bexpr)
return traj_objs
def _spawn_sco_var_for_pred(self, pred, t):
i = 0
x = np.empty(pred.x_dim, dtype=object)
v = np.empty(pred.x_dim)
for p in pred.attr_inds:
for attr, ind_arr in pred.attr_inds[p]:
n_vals = len(ind_arr)
ll_p = self._param_to_ll[p]
if p.is_symbol():
x[i:i + n_vals] = getattr(ll_p, attr)[ind_arr, 0]
v[i:i + n_vals] = getattr(p, attr)[ind_arr, 0]
else:
x[i:i + n_vals] = getattr(ll_p, attr)[ind_arr,
t - self.ll_start]
v[i:i + n_vals] = getattr(p, attr)[ind_arr, t]
i += n_vals
if pred.dynamic:
## include the variables from the next time step
for p in pred.attr_inds:
for attr, ind_arr in pred.attr_inds[p]:
n_vals = len(ind_arr)
ll_p = self._param_to_ll[p]
if p.is_symbol():
x[i:i + n_vals] = getattr(ll_p, attr)[ind_arr, 0]
v[i:i + n_vals] = getattr(p, attr)[ind_arr, 0]
else:
x[i:i + n_vals] = getattr(ll_p,
attr)[ind_arr,
t + 1 - self.ll_start]
v[i:i + n_vals] = getattr(p, attr)[ind_arr, t + 1]
i += n_vals
assert i >= pred.x_dim
x = x.reshape((pred.x_dim, 1))
v = v.reshape((pred.x_dim, 1))
return Variable(x, v)
def _solve_opt_prob(self,
plan,
priority,
callback=None,
active_ts=None,
verbose=False,
resample=True):
## active_ts is the inclusive timesteps to include
## in the optimization
if active_ts == None:
active_ts = (0, plan.horizon - 1)
model = grb.Model()
model.params.OutputFlag = 0
self._prob = Prob(model, callback=callback)
# param_to_ll_old = self._param_to_ll.copy()
self._spawn_parameter_to_ll_mapping(model, plan, active_ts)
model.update()
self._bexpr_to_pred = {}
if priority == -1:
obj_bexprs = self._get_trajopt_obj(plan, active_ts)
self._add_obj_bexprs(obj_bexprs)
self._add_first_and_last_timesteps_of_actions(plan,
priority=-1,
active_ts=active_ts,
verbose=verbose)
tol = 1e-2
elif priority == 0:
## this should only get called with a full plan for now
assert active_ts == (0, plan.horizon - 1)
obj_bexprs = []
if resample:
failed_preds = plan.get_failed_preds()
## this is an objective that places
## a high value on matching the resampled values
obj_bexprs.extend(self._resample(plan, failed_preds))
## solve an optimization movement primitive to
## transfer current trajectories
obj_bexprs.extend(self._get_transfer_obj(plan, self.transfer_norm))
self._add_obj_bexprs(obj_bexprs)
# self._add_first_and_last_timesteps_of_actions(
# plan, priority=0, add_nonlin=False)
self._add_first_and_last_timesteps_of_actions(plan,
priority=-1,
add_nonlin=True,
verbose=verbose)
self._add_all_timesteps_of_actions(plan,
priority=0,
add_nonlin=False,
verbose=verbose)
tol = 1e-2
elif priority == 1:
obj_bexprs = self._get_trajopt_obj(plan, active_ts)
self._add_obj_bexprs(obj_bexprs)
self._add_all_timesteps_of_actions(plan,
priority=1,
add_nonlin=True,
active_ts=active_ts,
verbose=verbose)
tol = 1e-4
solv = Solver()
solv.initial_penalty_coeff = self.init_penalty_coeff
success = solv.solve(self._prob,
method='penalty_sqp',
tol=tol,
verbose=verbose)
self._update_ll_params()
self._failed_groups = self._prob.nonconverged_groups
return success
class CanSolver(LLSolver):
def __init__(self, early_converge=False):
self.transfer_coeff = 1e1
self.rs_coeff = 1e10
self.initial_trust_region_size = 1e-2
self.init_penalty_coeff = 1e1
# self.init_penalty_coeff = 1e5
self.max_merit_coeff_increases = 5
self._param_to_ll = {}
self.early_converge = early_converge
self.child_solver = None
self.solve_priorities = [2]
def _solve_helper(self, plan, callback, active_ts, verbose):
# certain constraints should be solved first
success = False
for priority in self.solve_priorities:
success = self._solve_opt_prob(plan,
priority=priority,
callback=callback,
active_ts=active_ts,
verbose=verbose)
# if not success:
# return success
# return True
return success
def backtrack_solve(self, plan, callback=None, anum=0, verbose=False):
if anum > len(plan.actions) - 1:
return True
a = plan.actions[anum]
active_ts = a.active_timesteps
inits = {}
# Moveto -> Grasp -> Movetoholding -> Putdown
robot = a.params[0]
lookup_preds = 'EEReachable'
if a.name == 'moveto':
## moveto: (?robot - Robot ?start - RobotPose ?end - RobotPose)
## sample ?end - RobotPose
rs_param = a.params[2]
elif a.name == 'movetoholding':
## movetoholding: (?robot - Robot ?start - RobotPose ?end - RobotPose ?c - Can)
## sample ?end - RobotPose
rs_param = a.params[2]
elif a.name == 'grasp':
## grasp: (?robot - Robot ?can - Can ?target - Target ?sp - RobotPose ?ee - EEPose ?ep - RobotPose)
## sample ?ee - EEPose
rs_param = a.params[4]
lookup_preds = 'InContact'
elif a.name == 'putdown':
## putdown: (?robot - Robot ?can - Can ?target - Target ?sp - RobotPose ?ee - EEPose ?ep - RobotPose)
## sample ?ep - RobotPose
rs_param = a.params[5]
else:
raise NotImplemented
def recursive_solve():
## don't optimize over any params that are already set
old_params_free = {}
for p in plan.params.itervalues():
old_param_map = {}
if p.is_symbol():
if p not in a.params: continue
for attr in attr_map[getattr(p, '_type')]:
old_param_map[attr] = p._free_attrs[attr].copy()
p._free_attrs[attr][:] = 0
else:
for attr in attr_map[getattr(p, '_type')]:
old_param_map[attr] = p._free_attrs[
attr][:, active_ts[1]].copy()
p._free_attrs[attr][:, active_ts[1]] = 0
old_params_free[p] = old_param_map
self.child_solver = CanSolver()
if self.child_solver.backtrack_solve(plan,
callback=callback,
anum=anum + 1,
verbose=verbose):
return True
## reset free_attrs
for p in a.params:
if p.is_symbol():
if p not in a.params: continue
for attr in attr_map[getattr(p, '_type')]:
p._free_attrs[attr] = old_params_free[p][attr]
else:
for attr in attr_map[getattr(p, '_type')]:
p._free_attrs[attr][:, active_ts[1]] = old_params_free[
p][attr]
return False
if rs_param.is_fixed(attr_map[getattr(rs_param, '_type')]):
## this parameter is fixed
if callback is not None:
callback_a = lambda: callback(a)
else:
callback_a = None
self.child_solver = CanSolver()
success = self.child_solver.solve(plan,
callback=callback_a,
n_resamples=0,
active_ts=active_ts,
verbose=verbose,
force_init=True)
if not success:
## if planning fails we're done
return False
## no other options, so just return here
return recursive_solve()
## so that this won't be optimized over
rs_free = {}
for attr in attr_map[getattr(rs_param, '_type')]:
rs_free[attr] = rs_param._free_attrs[attr].copy()
rs_param._free_attrs[attr][:] = 0
if callback is not None:
callback_a = lambda: callback(a)
else:
callback_a = None
######################################################################
# Want robot_poses to be a list of dict, mapping from attr to value
robot_poses = []
if a.name == 'moveto': #rs_param is ?end - RobotPose
ee_poses = plan.get_param(lookup_preds, 2, {1: rs_param})
# get a collision free base_pose, with all other pose stay the same
elif a.name == 'movetoholding': #rs_param is ?end - RobotPose
ee_poses = plan.get_param(lookup_preds, 2, {1: rs_param})
targets = plan.get_param('InContact', 2, {1: ee_poses})
# sample ee_pose associated to this target and sample robot_pose for it
elif a.name == 'grasp': #rs_param is ?ee - EEPose, robot_pose
ee_poses = [rs_param]
targets = plan.get_param(lookup_preds, 2, {1: rs_param})
# sample ee_pose associated to this target and sample robot_pose for it
elif a.name == 'putdown':
ee_poses = plan.get_param(lookup_preds, 2, {1: rs_param})
targets = []
# get a collision free base_pose, with all other pose stay the same
else:
raise NotImplemented
# import ipdb; ipdb.set_trace()
if len(ee_poses) > 1:
import pdb
pdb.set_trace()
if len(ee_poses) == 0 or np.all([
ee_poses[0]._free_attrs[attr]
for attr in attr_map[ee_poses[0]._type]
]):
# Currently just sample reasonable robot end poses for the moveto action
t = active_ts[0]
ee_pose = ee_poses[0]
targets = plan.get_param('InContact', 2, {1: ee_pose})
if len(targets) == 0 or len(targets) > 1:
import ipdb
ipdb.set_trace()
sampled_ee_poses = sampling.get_ee_from_target(
targets[0].value, targets[0].rotation)
cur_robot_pose = {
'backHeight': robot.backHeight[:, t],
'lArmPose': robot.lArmPose[:, t],
'lGripper': robot.lGripper[:, t],
'rArmPose': robot.rArmPose[:, t],
'rGripper': robot.rGripper[:, t]
}
for ee_pose in sampled_ee_poses:
possible_base = sampling.get_base_poses_around_pos(
t, robot, ee_pose[0], BASE_SAMPLE_SIZE, 0.6)
robot_base = sampling.closest_base_poses(
possible_base, robot.pose[:, t])
robot_pose = cur_robot_pose.copy()
robot_pose['value'] = robot_base
robot_poses.append(robot_pose)
elif np.any([
ee_poses[0]._free_attrs[attr]
for attr in attr_map[ee_poses[0]._type]
]):
## there shouldn't be only some free_attrs set
raise NotImplementedError
else:
ee_pose = ee_poses[0]
targets = plan.get_param('InContact', 2, {1: ee_pose})
possible_ee_poses = sampling.get_ee_from_target(
targets[0].value, targets[0].rotation)
for ee_pos, ee_rot in possible_ee_poses:
robot_poses.append({'value': ee_pos, 'rotation': ee_rot})
# import ipdb; ipdb.set_trace()
##########################################################################
for rp in robot_poses:
for attr in attr_map[rs_param._type]:
dim = len(rp[attr])
setattr(rs_param, attr, rp[attr].reshape((dim, 1)))
success = False
self.child_solver = CanSolver()
success = self.child_solver.solve(plan,
callback=callback_a,
n_resamples=0,
active_ts=active_ts,
verbose=verbose,
force_init=True)
if success:
if recursive_solve():
break
else:
success = False
for attr in attr_map[getattr(rs_param, '_type')]:
rs_param._free_attrs[attr] = rs_free[attr]
return success
def sample_rp_from_ee(t, plan, ee_poses, robot):
target_pose = target.value[:, 0]
dummy_body = OpenRAVEBody(plan.env, robot.name + '_dummy', robot.geom)
possible_rps = []
for ee_pose in ee_poses:
base_pose = sampling.get_col_free_base_pose_around_target(
t, plan, ee_pose.value, robot).reshape((1, 3))
ik_arm_poses = dummy_body.ik_arm_pose(ee_pose[0], ee_pose[1])
cur_arm_poses = np.c_[robot.backHeight, robot.rArmPose].reshape(
(8, ))
chosen_pose = sampling.closest_arm_pose(ik_arm_poses,
cur_arm_poses)
torso_pose, arm_pose = chosen_pose[0], chosen_pose[1:]
possible_rp = np.hstack(
(base_pose, torso_pose, robot.lArmPose[:, t],
robot.lGripper[:, t], rArmPose, robot.rGripper[:, t]))
possible_rps.append(possible_rp)
dummy_body.delete()
if len(possible_rps) <= 0:
print "something went wrong"
# import ipdb; ipdb.set_trace()
return possible_rps
def solve(self,
plan,
callback=None,
n_resamples=5,
active_ts=None,
verbose=False,
force_init=False):
success = False
if force_init or not plan.initialized:
## solve at priority -1 to get an initial value for the parameters
self._solve_opt_prob(plan,
priority=-2,
callback=callback,
active_ts=active_ts,
verbose=verbose)
self._solve_opt_prob(plan,
priority=-1,
callback=callback,
active_ts=active_ts,
verbose=verbose)
plan.initialized = True
success = self._solve_helper(plan,
callback=callback,
active_ts=active_ts,
verbose=verbose)
# fp = plan.get_failed_preds()
# if len(fp) == 0:
# return True
if success:
return success
for _ in range(n_resamples):
## refinement loop
## priority 0 resamples the first failed predicate in the plan
## and then solves a transfer optimization that only includes linear constraints
self._solve_opt_prob(plan,
priority=0,
callback=callback,
active_ts=active_ts,
verbose=verbose)
# self._solve_opt_prob(plan, priority=1, callback=callback, active_ts=active_ts, verbose=verbose)
success = self._solve_opt_prob(plan,
priority=2,
callback=callback,
active_ts=active_ts,
verbose=verbose)
# fp = plan.get_failed_preds()
# if len(fp) == 0:
# return True
if success:
return success
return success
def _solve_opt_prob(self,
plan,
priority,
callback=None,
init=True,
active_ts=None,
verbose=False):
## active_ts is the inclusive timesteps to include
## in the optimization
if active_ts == None:
active_ts = (0, plan.horizon - 1)
model = grb.Model()
model.params.OutputFlag = 0
self._prob = Prob(model, callback=callback)
# param_to_ll_old = self._param_to_ll.copy()
self._spawn_parameter_to_ll_mapping(model, plan, active_ts)
model.update()
self._bexpr_to_pred = {}
if priority == -2:
obj_bexprs = self._get_trajopt_obj(plan, active_ts)
self._add_obj_bexprs(obj_bexprs)
self._add_first_and_last_timesteps_of_actions(
plan,
priority=MAX_PRIORITY,
active_ts=active_ts,
verbose=verbose,
add_nonlin=False)
# self._add_all_timesteps_of_actions(plan, priority=1, add_nonlin=False, verbose=verbose)
tol = 1e-1
elif priority == -1:
obj_bexprs = self._get_trajopt_obj(plan, active_ts)
self._add_obj_bexprs(obj_bexprs)
self._add_first_and_last_timesteps_of_actions(
plan,
priority=MAX_PRIORITY,
active_ts=active_ts,
verbose=verbose,
add_nonlin=True)
tol = 1e-1
elif priority == 0:
## this should only get called with a full plan for now
assert active_ts == (0, plan.horizon - 1)
failed_preds = plan.get_failed_preds()
## this is an objective that places
## a high value on matching the resampled values
obj_bexprs = []
obj_bexprs.extend(self._resample(plan, failed_preds))
## solve an optimization movement primitive to
## transfer current trajectories
obj_bexprs.extend(self._get_trajopt_obj(plan, active_ts))
# obj_bexprs.extend(self._get_transfer_obj(plan, 'min-vel'))
self._add_obj_bexprs(obj_bexprs)
# self._add_first_and_last_timesteps_of_actions(
# plan, priority=0, add_nonlin=False)
# self._add_first_and_last_timesteps_of_actions(
# plan, priority=-1, add_nonlin=True, verbose=verbose)
# self._add_all_timesteps_of_actions(
# plan, priority=0, add_nonlin=False, verbose=verbose)
self._add_all_timesteps_of_actions(plan,
priority=1,
add_nonlin=True,
verbose=verbose)
# self._add_first_and_last_timesteps_of_actions(
# plan, priority=1, add_nonlin=True, verbose=verbose)
# self._add_all_timesteps_of_actions(
# plan, priority=0, add_nonlin=False, verbose=verbose)
tol = 1e-1
elif priority >= 1:
obj_bexprs = self._get_trajopt_obj(plan, active_ts)
self._add_obj_bexprs(obj_bexprs)
self._add_all_timesteps_of_actions(plan,
priority=priority,
add_nonlin=True,
active_ts=active_ts,
verbose=verbose)
tol = 1e-3
solv = Solver()
solv.initial_trust_region_size = self.initial_trust_region_size
solv.initial_penalty_coeff = self.init_penalty_coeff
solv.max_merit_coeff_increases = self.max_merit_coeff_increases
success = solv.solve(self._prob,
method='penalty_sqp',
tol=tol,
verbose=True)
self._update_ll_params()
print "priority: {}".format(priority)
# if callback is not None: callback(True)
return success
def _get_transfer_obj(self, plan, norm):
transfer_objs = []
if norm == 'min-vel':
for param in plan.params.values():
# if param._type in ['Robot', 'Can', 'EEPose']:
for attr_name in param.__dict__.iterkeys():
attr_type = param.get_attr_type(attr_name)
if issubclass(attr_type, Vector):
if param.is_symbol():
T = 1
else:
T = plan.horizon
K = attr_type.dim
attr_val = getattr(param, attr_name)
# pose = param.pose
assert (K, T) == attr_val.shape
KT = K * T
v = -1 * np.ones((KT - K, 1))
d = np.vstack((np.ones((KT - K, 1)), np.zeros((K, 1))))
# [:,0] allows numpy to see v and d as one-dimensional so
# that numpy will create a diagonal matrix with v and d as a diagonal
P = np.diag(v[:, 0], K) + np.diag(d[:, 0])
Q = np.dot(np.transpose(P), P)
cur_val = attr_val.reshape((KT, 1), order='F')
A = -2 * cur_val.T.dot(Q)
b = cur_val.T.dot(Q.dot(cur_val))
# QuadExpr is 0.5*x^Tx + Ax + b
quad_expr = QuadExpr(2 * self.transfer_coeff * Q,
self.transfer_coeff * A,
self.transfer_coeff * b)
param_ll = self._param_to_ll[param]
ll_attr_val = getattr(param_ll, attr_name)
param_ll_grb_vars = ll_attr_val.reshape((KT, 1),
order='F')
bexpr = BoundExpr(quad_expr,
Variable(param_ll_grb_vars, cur_val))
transfer_objs.append(bexpr)
else:
raise NotImplemented
return transfer_objs
def _resample(self, plan, preds):
val, attr_inds = None, None
for negated, pred, t in preds:
## returns a vector of new values and an
## attr_inds (OrderedDict) that gives the mapping
## to parameter attributes
val, attr_inds = pred.resample(negated, t, plan)
## no resample defined for that pred
if val is not None: break
if val is None:
return []
bexprs = []
i = 0
for p in attr_inds:
## get the ll_param for p and gurobi variables
ll_p = self._param_to_ll[p]
n_vals = 0
grb_vars = []
for attr, ind_arr, t in attr_inds[p]:
n_vals += len(ind_arr)
grb_vars.extend(list(
getattr(ll_p, attr)[ind_arr, t].flatten()))
for j, grb_var in enumerate(grb_vars):
## create an objective saying stay close to this value
## e(x) = x^2 - 2*val[i+j]*x + val[i+j]^2
Q = np.eye(1)
A = -2 * val[i + j] * np.ones((1, 1))
b = np.ones((1, 1)) * np.power(val[i + j], 2)
# QuadExpr is 0.5*x^Tx + Ax + b
quad_expr = QuadExpr(2 * Q * self.rs_coeff, A * self.rs_coeff,
b * self.rs_coeff)
v_arr = np.array([grb_var]).reshape((1, 1), order='F')
init_val = np.ones((1, 1)) * val[i + j]
bexpr = BoundExpr(quad_expr,
Variable(v_arr, val[i + j].reshape((1, 1))))
bexprs.append(bexpr)
i += n_vals
return bexprs
def _add_pred_dict(self,
pred_dict,
effective_timesteps,
add_nonlin=True,
priority=MAX_PRIORITY,
verbose=False):
## for debugging
ignore_preds = []
priority = np.maximum(priority, 0)
if not pred_dict['hl_info'] == "hl_state":
start, end = pred_dict['active_timesteps']
active_range = range(start, end + 1)
if verbose:
print "pred being added: ", pred_dict
print active_range, effective_timesteps
negated = pred_dict['negated']
pred = pred_dict['pred']
if pred.get_type() in ignore_preds:
return
if pred.priority > priority: return
assert isinstance(pred, common_predicates.ExprPredicate)
expr = pred.get_expr(negated)
for t in effective_timesteps:
if t in active_range:
if expr is not None:
if add_nonlin or isinstance(expr.expr, AffExpr):
if verbose:
print "expr being added at time ", t
var = self._spawn_sco_var_for_pred(pred, t)
bexpr = BoundExpr(expr, var)
# TODO: REMOVE line below, for tracing back predicate for debugging.
bexpr.pred = pred
self._bexpr_to_pred[bexpr] = (negated, pred, t)
groups = ['all']
if self.early_converge:
## this will check for convergence per parameter
## this is good if e.g., a single trajectory quickly
## gets stuck
groups.extend(
[param.name for param in pred.params])
self._prob.add_cnt_expr(bexpr, groups)
def _add_first_and_last_timesteps_of_actions(self,
plan,
priority=MAX_PRIORITY,
add_nonlin=False,
active_ts=None,
verbose=False):
if active_ts == None:
active_ts = (0, plan.horizon - 1)
for action in plan.actions:
action_start, action_end = action.active_timesteps
## only add an action
if action_start >= active_ts[1] and action_start > active_ts[0]:
continue
if action_end < active_ts[0]: continue
for pred_dict in action.preds:
if action_start >= active_ts[0]:
self._add_pred_dict(pred_dict, [action_start],
priority=priority,
add_nonlin=add_nonlin,
verbose=verbose)
if action_end <= active_ts[1]:
self._add_pred_dict(pred_dict, [action_end],
priority=priority,
add_nonlin=add_nonlin,
verbose=verbose)
## add all of the linear ineqs
timesteps = range(max(action_start + 1, active_ts[0]),
min(action_end, active_ts[1]))
for pred_dict in action.preds:
self._add_pred_dict(pred_dict,
timesteps,
add_nonlin=False,
priority=priority,
verbose=verbose)
def _add_all_timesteps_of_actions(self,
plan,
priority=MAX_PRIORITY,
add_nonlin=True,
active_ts=None,
verbose=False):
if active_ts == None:
active_ts = (0, plan.horizon - 1)
for action in plan.actions:
action_start, action_end = action.active_timesteps
if action_start >= active_ts[1] and action_start > active_ts[0]:
continue
if action_end < active_ts[0]: continue
timesteps = range(max(action_start, active_ts[0]),
min(action_end, active_ts[1]) + 1)
for pred_dict in action.preds:
self._add_pred_dict(pred_dict,
timesteps,
priority=priority,
add_nonlin=add_nonlin,
verbose=verbose)
def _update_ll_params(self):
for ll_param in self._param_to_ll.values():
ll_param.update_param()
if self.child_solver:
self.child_solver._update_ll_params()
def _spawn_parameter_to_ll_mapping(self, model, plan, active_ts=None):
if active_ts == None:
active_ts = (0, plan.horizon - 1)
horizon = active_ts[1] - active_ts[0] + 1
self._param_to_ll = {}
self.ll_start = active_ts[0]
for param in plan.params.values():
ll_param = LLParam(model, param, horizon, active_ts)
ll_param.create_grb_vars()
self._param_to_ll[param] = ll_param
model.update()
for ll_param in self._param_to_ll.values():
ll_param.batch_add_cnts()
def _add_obj_bexprs(self, obj_bexprs):
for bexpr in obj_bexprs:
self._prob.add_obj_expr(bexpr)
def _get_trajopt_obj(self, plan, active_ts=None):
if active_ts == None:
active_ts = (0, plan.horizon - 1)
start, end = active_ts
traj_objs = []
for param in plan.params.values():
if param not in self._param_to_ll:
continue
if param._type in ['Robot', 'Can']:
for attr_name in param.__dict__.iterkeys():
attr_type = param.get_attr_type(attr_name)
if issubclass(attr_type, Vector):
T = end - start + 1
K = attr_type.dim
attr_val = getattr(param, attr_name)
KT = K * T
v = -1 * np.ones((KT - K, 1))
d = np.vstack((np.ones((KT - K, 1)), np.zeros((K, 1))))
# [:,0] allows numpy to see v and d as one-dimensional so
# that numpy will create a diagonal matrix with v and d as a diagonal
P = np.diag(v[:, 0], K) + np.diag(d[:, 0])
Q = np.dot(np.transpose(P), P)
Q *= TRAJOPT_COEFF
quad_expr = None
if attr_name == 'pose' and param._type == 'Robot':
quad_expr = QuadExpr(BASE_MOVE_COEFF * Q,
np.zeros((1, KT)),
np.zeros((1, 1)))
else:
quad_expr = QuadExpr(Q, np.zeros((1, KT)),
np.zeros((1, 1)))
param_ll = self._param_to_ll[param]
ll_attr_val = getattr(param_ll, attr_name)
param_ll_grb_vars = ll_attr_val.reshape((KT, 1),
order='F')
attr_val = getattr(param, attr_name)
init_val = attr_val[:, start:end + 1].reshape(
(KT, 1), order='F')
bexpr = BoundExpr(
quad_expr, Variable(param_ll_grb_vars, init_val))
# bexpr = BoundExpr(quad_expr, Variable(param_ll_grb_vars))
traj_objs.append(bexpr)
return traj_objs
def _solve_opt_prob(self,
plan,
priority,
callback=None,
init=True,
active_ts=None,
verbose=False):
## active_ts is the inclusive timesteps to include
## in the optimization
if active_ts == None:
active_ts = (0, plan.horizon - 1)
model = grb.Model()
model.params.OutputFlag = 0
self._prob = Prob(model, callback=callback)
# param_to_ll_old = self._param_to_ll.copy()
self._spawn_parameter_to_ll_mapping(model, plan, active_ts)
model.update()
self._bexpr_to_pred = {}
if priority == -2:
obj_bexprs = self._get_trajopt_obj(plan, active_ts)
self._add_obj_bexprs(obj_bexprs)
self._add_first_and_last_timesteps_of_actions(
plan,
priority=MAX_PRIORITY,
active_ts=active_ts,
verbose=verbose,
add_nonlin=False)
# self._add_all_timesteps_of_actions(plan, priority=1, add_nonlin=False, verbose=verbose)
tol = 1e-1
elif priority == -1:
obj_bexprs = self._get_trajopt_obj(plan, active_ts)
self._add_obj_bexprs(obj_bexprs)
self._add_first_and_last_timesteps_of_actions(
plan,
priority=MAX_PRIORITY,
active_ts=active_ts,
verbose=verbose,
add_nonlin=True)
tol = 1e-1
elif priority == 0:
## this should only get called with a full plan for now
assert active_ts == (0, plan.horizon - 1)
failed_preds = plan.get_failed_preds()
## this is an objective that places
## a high value on matching the resampled values
obj_bexprs = []
obj_bexprs.extend(self._resample(plan, failed_preds))
## solve an optimization movement primitive to
## transfer current trajectories
obj_bexprs.extend(self._get_trajopt_obj(plan, active_ts))
# obj_bexprs.extend(self._get_transfer_obj(plan, 'min-vel'))
self._add_obj_bexprs(obj_bexprs)
# self._add_first_and_last_timesteps_of_actions(
# plan, priority=0, add_nonlin=False)
# self._add_first_and_last_timesteps_of_actions(
# plan, priority=-1, add_nonlin=True, verbose=verbose)
# self._add_all_timesteps_of_actions(
# plan, priority=0, add_nonlin=False, verbose=verbose)
self._add_all_timesteps_of_actions(plan,
priority=1,
add_nonlin=True,
verbose=verbose)
# self._add_first_and_last_timesteps_of_actions(
# plan, priority=1, add_nonlin=True, verbose=verbose)
# self._add_all_timesteps_of_actions(
# plan, priority=0, add_nonlin=False, verbose=verbose)
tol = 1e-1
elif priority >= 1:
obj_bexprs = self._get_trajopt_obj(plan, active_ts)
self._add_obj_bexprs(obj_bexprs)
self._add_all_timesteps_of_actions(plan,
priority=priority,
add_nonlin=True,
active_ts=active_ts,
verbose=verbose)
tol = 1e-3
solv = Solver()
solv.initial_trust_region_size = self.initial_trust_region_size
solv.initial_penalty_coeff = self.init_penalty_coeff
solv.max_merit_coeff_increases = self.max_merit_coeff_increases
success = solv.solve(self._prob,
method='penalty_sqp',
tol=tol,
verbose=True)
self._update_ll_params()
print "priority: {}".format(priority)
# if callback is not None: callback(True)
return success
class CanSolver(LLSolver):
def __init__(self, early_converge=False):
self.transfer_coeff = 1e1
self.rs_coeff = 1e6
self.initial_trust_region_size = 1e-1
self.init_penalty_coeff = 1e1
# self.init_penalty_coeff = 1e5
self.max_merit_coeff_increases = 5
self._param_to_ll = {}
self.early_converge = early_converge
self.child_solver = None
def backtrack_solve(self, plan, callback=None, anum=0, verbose=False):
if anum > len(plan.actions) - 1:
return True
a = plan.actions[anum]
active_ts = a.active_timesteps
inits = {}
# Moveto -> Grasp -> Movetoholding -> Putdown
if a.name == 'moveto':
## moveto: (?robot - Robot ?start - RobotPose ?end - RobotPose)
## find possible values for the final pose
rs_param = a.params[2]
elif a.name == 'movetoholding':
## movetoholding: (?robot - Robot ?start - RobotPose ?end - RobotPose ?c - Can)
## find possible values for the final pose
rs_param = a.params[3]
elif a.name == 'grasp':
## grasp: (?robot - Robot ?can - Can ?target - Target ?sp - RobotPose ?ee - EEPose ?ep - RobotPose)
## sample the ee_pose
rs_param = a.params[5]
elif a.name == 'putdown':
## putdown: (?robot - Robot ?can - Can ?target - Target ?sp - RobotPose ?ee - EEPose ?ep - RobotPose)
rs_param = a.params[4]
else:
raise NotImplemented
def recursive_solve():
## don't optimize over any params that are already set
old_params_free = {}
for p in plan.params.itervalues():
old_param_map = {}
if p.is_symbol():
if p not in a.params: continue
for attr in attr_map[getattr(p, '_type')]:
old_param_map[attr] = p._free_attrs[attr].copy()
p._free_attrs[attr][:] = 0
else:
for attr in attr_map[getattr(p, '_type')]:
old_param_map[attr] = p._free_attrs[
attr][:, active_ts[1]].copy()
p._free_attrs[attr][:, active_ts[1]] = 0
old_params_free[p] = old_param_map
self.child_solver = CanSolver()
if self.child_solver.backtrack_solve(plan,
callback=callback,
anum=anum + 1,
verbose=verbose):
return True
## reset free_attrs
for p in a.params:
if p.is_symbol():
if p not in a.params: continue
for attr in attr_map[getattr(p, '_type')]:
p._free_attrs[attr] = old_params_free[p][attr]
else:
for attr in attr_map[getattr(p, '_type')]:
p._free_attrs[attr][:, active_ts[1]] = old_params_free[
p][attr]
return False
if rs_param.is_fixed(attr_map[getattr(rs_param, '_type')]):
## this parameter is fixed
if callback is not None:
callback_a = lambda: callback(a)
else:
callback_a = None
self.child_solver = CanSolver()
success = self.child_solver.solve(plan,
callback=callback_a,
n_resamples=0,
active_ts=active_ts,
verbose=verbose,
force_init=True)
if not success:
## if planning fails we're done
return False
## no other options, so just return here
return recursive_solve()
## so that this won't be optimized over
rs_free = {}
for attr in attr_map[getattr(rs_param, '_type')]:
rs_free[attr] = rs_param._free_attrs[attr].copy()
rs_param._free_attrs[attr][:] = 0
# Target is needed to sample the ee_pose, ee_pose is needed to sample the robot_pose
# InContact, Robot, EEPose, Target
targets = plan.get_param('InContact', 2, {1: rs_param}, negated=False)
# EEReachable, Robot, RobotPose, EEPose
robot_pose = plan.get_param('EEReachable',
1, {2: rs_param},
negated=False)
if len(targets) > 1:
import pdb
pdb.set_trace()
if callback is not None:
callback_a = lambda: callback(a)
else:
callback_a = None
robot_poses = []
if len(targets) == 0 or np.all(
[targets[0]._free_attrs[attr] for attr in attr_map[targets[0]]]):
## sample 4 possible poses
coords = list(itertools.product(range(WIDTH), range(HEIGHT)))
random.shuffle(coords)
robot_poses = [np.array(x)[:, None] for x in coords[:4]]
elif np.any(
[targets[0]._free_attrs[attr] for attr in attr_map[targets[0]]]):
## there shouldn't be only some free_attrs set
raise NotImplementedError
else:
grasp_dirs = [
np.array([0, -1]),
np.array([1, 0]),
np.array([0, 1]),
np.array([-1, 0])
]
grasp_len = plan.params['pr2'].geom.radius + targets[0].geom.radius
for g_dir in grasp_dirs:
grasp = (g_dir * grasp_len).reshape((2, 1))
robot_poses.append(targets[0].value + grasp)
for rp in robot_poses:
rs_param.value = rp
success = False
self.child_solver = CanSolver()
success = self.child_solver.solve(plan,
callback=callback_a,
n_resamples=0,
active_ts=active_ts,
verbose=verbose,
force_init=True)
if success:
if recursive_solve():
break
else:
success = False
for attr in attr_map[getattr(rs_param, '_type')]:
rs_param._free_attrs[attr] = rs_free[attr]
return success
def sample_ee_from_target(self, target):
"""
Sample all possible EE Pose that pr2 can grasp with
target: parameter of type Target
return: list of tuple in the format of (ee_pos, ee_rot)
"""
possible_ee_poses = []
targ_pos, targ_rot = target.value.flatten(), target.rotation.flatten()
ee_pos = targ_pos.copy()
target_trans = OpenRAVEBody.transform_from_obj_pose(targ_pos, targ_rot)
# rotate can's local z-axis by the amount of linear spacing between 0 to 2pi
angle_range = np.linspace(0, np.pi * 2, num=SAMPLE_SIZE)
for rot in angle_range:
target_trans = OpenRAVEBody.transform_from_obj_pose(
targ_pos, targ_rot)
# rotate new ee_pose around can's rotation axis
rot_mat = matrixFromAxisAngle([0, 0, rot])
ee_trans = target_trans.dot(rot_mat)
ee_rot = OpenRAVEBody.obj_pose_from_transform(ee_trans)[3:]
possible_ee_poses.append((ee_pos, ee_rot))
return possible_ee_poses
def sample_rp_from_ee(self, env, rs_param, ee_pose):
"""
Using Inverse Kinematics to solve all possible robot armpose and basepose that reaches this ee_pose
robot_pose: list of full robot pose with format of(basepose, backHeight, lArmPose, lGripper, rArmPose, rGripper)
"""
ee_pos, ee_rot = ee_pose[0], ee_pose[1]
base_raidus = np.linspace(-1, 1, num=BASE_SAMPLE_SIZE)
poses = list(itertools.product(base_raidus, base_raidus))
# sample
dummy_body = OpenRAVEBody(env, rs_param.name, PR2())
possible_robot_pose = []
for pos in poses:
bp = pos + ee_pos[:2]
vec = ee_pos[:2] - bp
vec = vec / np.linalg.norm(vec)
theta = math.atan2(vec[1], vec[0])
bp = np.array([bp[0], bp[1], theta])
dummy_body.set_pose(bp)
arm_poses = robot_body.ik_arm_pose(ee_pos, ee_rot)
for arm_pose in arm_poses:
possible_robot_pose.append(
(bp, ee_arm_poses[0], rs_param.lArmPose, rs_param.lGripper,
ee_arm_poses[1:], rs_param.rGripper))
import ipdb
ipdb.set_trace()
dummy_body.delete()
return possible_robot_pose
def closest_arm_pose(self, arm_poses, cur_arm_poses):
pass
def solve(self,
plan,
callback=None,
n_resamples=5,
active_ts=None,
verbose=False,
force_init=False):
success = False
if force_init or not plan.initialized:
## solve at priority -1 to get an initial value for the parameters
self._solve_opt_prob(plan,
priority=-1,
callback=callback,
active_ts=active_ts,
verbose=verbose)
plan.initialized = True
success = self._solve_opt_prob(plan,
priority=1,
callback=callback,
active_ts=active_ts,
verbose=verbose)
if success:
return success
for _ in range(n_resamples):
## refinement loop
## priority 0 resamples the first failed predicate in the plan
## and then solves a transfer optimization that only includes linear constraints
self._solve_opt_prob(plan,
priority=0,
callback=callback,
active_ts=active_ts,
verbose=verbose)
success = self._solve_opt_prob(plan,
priority=1,
callback=callback,
active_ts=active_ts,
verbose=verbose)
if success:
return success
return success
def _solve_opt_prob(self,
plan,
priority,
callback=None,
init=True,
init_from_prev=False,
active_ts=None,
verbose=False):
## active_ts is the inclusive timesteps to include
## in the optimization
if active_ts == None:
active_ts = (0, plan.horizon - 1)
model = grb.Model()
model.params.OutputFlag = 0
self._prob = Prob(model, callback=callback)
# param_to_ll_old = self._param_to_ll.copy()
self._spawn_parameter_to_ll_mapping(model, plan, active_ts)
model.update()
self._bexpr_to_pred = {}
if priority == -1:
obj_bexprs = self._get_trajopt_obj(plan, active_ts)
self._add_obj_bexprs(obj_bexprs)
self._add_first_and_last_timesteps_of_actions(plan,
priority=-1,
active_ts=active_ts,
verbose=verbose)
tol = 1e-1
elif priority == 0:
## this should only get called with a full plan for now
assert active_ts == (0, plan.horizon - 1)
failed_preds = plan.get_failed_preds()
## this is an objective that places
## a high value on matching the resampled values
obj_bexprs = self._resample(plan, failed_preds)
## solve an optimization movement primitive to
## transfer current trajectories
obj_bexprs.extend(self._get_transfer_obj(plan, 'min-vel'))
self._add_obj_bexprs(obj_bexprs)
# self._add_first_and_last_timesteps_of_actions(
# plan, priority=0, add_nonlin=False)
self._add_first_and_last_timesteps_of_actions(plan,
priority=-1,
add_nonlin=True,
verbose=verbose)
self._add_all_timesteps_of_actions(plan,
priority=0,
add_nonlin=False,
verbose=verbose)
tol = 1e-1
elif priority == 1:
obj_bexprs = self._get_trajopt_obj(plan, active_ts)
self._add_obj_bexprs(obj_bexprs)
self._add_all_timesteps_of_actions(plan,
priority=1,
add_nonlin=True,
active_ts=active_ts,
verbose=verbose)
tol = 1e-4
solv = Solver()
solv.initial_trust_region_size = self.initial_trust_region_size
solv.initial_penalty_coeff = self.init_penalty_coeff
solv.max_merit_coeff_increases = self.max_merit_coeff_increases
success = solv.solve(self._prob,
method='penalty_sqp',
tol=tol,
verbose=True)
import ipdb
ipdb.set_trace()
self._update_ll_params()
return success
def _get_transfer_obj(self, plan, norm):
transfer_objs = []
if norm == 'min-vel':
for param in plan.params.values():
if param._type in ['Robot', 'Can']:
T = plan.horizon
K = 2
pose = param.pose
assert (K, T) == pose.shape
KT = K * T
v = -1 * np.ones((KT - K, 1))
d = np.vstack((np.ones((KT - K, 1)), np.zeros((K, 1))))
# [:,0] allows numpy to see v and d as one-dimensional so
# that numpy will create a diagonal matrix with v and d as a diagonal
P = np.diag(v[:, 0], K) + np.diag(d[:, 0])
Q = np.dot(np.transpose(P), P)
cur_pose = pose.reshape((KT, 1), order='F')
A = -2 * cur_pose.T.dot(Q)
b = cur_pose.T.dot(Q.dot(cur_pose))
# QuadExpr is 0.5*x^Tx + Ax + b
quad_expr = QuadExpr(2 * self.transfer_coeff * Q,
self.transfer_coeff * A,
self.transfer_coeff * b)
ll_param = self._param_to_ll[param]
ll_grb_vars = ll_param.pose.reshape((KT, 1), order='F')
bexpr = BoundExpr(quad_expr,
Variable(ll_grb_vars, cur_pose))
transfer_objs.append(bexpr)
else:
raise NotImplemented
return transfer_objs
def _resample(self, plan, preds):
val, attr_inds = None, None
for negated, pred, t in preds:
## returns a vector of new values and an
## attr_inds (OrderedDict) that gives the mapping
## to parameter attributes
val, attr_inds = pred.resample(negated, t, plan)
## no resample defined for that pred
if val is not None: break
if val is None:
return None
t_local = t
bexprs = []
i = 0
for p in attr_inds:
## get the ll_param for p and gurobi variables
ll_p = self._param_to_ll[p]
if p.is_symbol(): t_local = 0
n_vals = 0
grb_vars = []
for attr, ind_arr in attr_inds[p]:
n_vals += len(ind_arr)
grb_vars.extend(list(getattr(ll_p, attr)[ind_arr, t_local]))
for j, grb_var in enumerate(grb_vars):
## create an objective saying stay close to this value
## e(x) = x^2 - 2*val[i+j]*x + val[i+j]^2
Q = np.eye(1)
A = -2 * val[i + j] * np.ones((1, 1))
b = np.ones((1, 1)) * np.power(val[i + j], 2)
# QuadExpr is 0.5*x^Tx + Ax + b
quad_expr = QuadExpr(2 * Q * self.rs_coeff, A * self.rs_coeff,
b * self.rs_coeff)
v_arr = np.array([grb_var]).reshape((1, 1), order='F')
init_val = np.ones((1, 1)) * val[i + j]
bexpr = BoundExpr(quad_expr,
Variable(v_arr, val[i + j].reshape((1, 1))))
bexprs.append(bexpr)
i += n_vals
return bexprs
def _add_pred_dict(self,
pred_dict,
effective_timesteps,
add_nonlin=True,
priority=MAX_PRIORITY,
verbose=False):
## for debugging
ignore_preds = []
priority = np.maximum(priority, 0)
if not pred_dict['hl_info'] == "hl_state":
start, end = pred_dict['active_timesteps']
active_range = range(start, end + 1)
if verbose:
print "pred being added: ", pred_dict
print active_range, effective_timesteps
negated = pred_dict['negated']
pred = pred_dict['pred']
if pred.get_type() in ignore_preds:
return
if pred.priority > priority: return
assert isinstance(pred, common_predicates.ExprPredicate)
expr = pred.get_expr(negated)
for t in effective_timesteps:
if t in active_range:
if expr is not None:
if add_nonlin or isinstance(expr.expr, AffExpr):
if verbose:
print "expr being added at time ", t
var = self._spawn_sco_var_for_pred(pred, t)
bexpr = BoundExpr(expr, var)
# TODO: REMOVE line below, for tracing back predicate for debugging.
bexpr.pred = pred
self._bexpr_to_pred[bexpr] = (negated, pred, t)
groups = ['all']
if self.early_converge:
## this will check for convergence per parameter
## this is good if e.g., a single trajectory quickly
## gets stuck
groups.extend(
[param.name for param in pred.params])
self._prob.add_cnt_expr(bexpr, groups)
def _add_first_and_last_timesteps_of_actions(self,
plan,
priority=MAX_PRIORITY,
add_nonlin=False,
active_ts=None,
verbose=False):
if active_ts == None:
active_ts = (0, plan.horizon - 1)
for action in plan.actions:
action_start, action_end = action.active_timesteps
## only add an action
if action_start >= active_ts[1]: continue
if action_end < active_ts[0]: continue
for pred_dict in action.preds:
if action_start >= active_ts[0]:
self._add_pred_dict(pred_dict, [action_start],
priority=priority,
add_nonlin=add_nonlin,
verbose=verbose)
if action_end <= active_ts[1]:
self._add_pred_dict(pred_dict, [action_end],
priority=priority,
add_nonlin=add_nonlin,
verbose=verbose)
## add all of the linear ineqs
timesteps = range(max(action_start + 1, active_ts[0]),
min(action_end, active_ts[1]))
for pred_dict in action.preds:
self._add_pred_dict(pred_dict,
timesteps,
add_nonlin=False,
priority=priority,
verbose=verbose)
def _add_all_timesteps_of_actions(self,
plan,
priority=MAX_PRIORITY,
add_nonlin=True,
active_ts=None,
verbose=False):
if active_ts == None:
active_ts = (0, plan.horizon - 1)
for action in plan.actions:
action_start, action_end = action.active_timesteps
if action_start > active_ts[1]: continue
if action_end < active_ts[0]: continue
timesteps = range(max(action_start, active_ts[0]),
min(action_end, active_ts[1]) + 1)
for pred_dict in action.preds:
self._add_pred_dict(pred_dict,
timesteps,
priority=priority,
add_nonlin=add_nonlin,
verbose=verbose)
def _update_ll_params(self):
for ll_param in self._param_to_ll.values():
ll_param.update_param()
if self.child_solver:
self.child_solver._update_ll_params()
def _spawn_parameter_to_ll_mapping(self, model, plan, active_ts=None):
if active_ts == None:
active_ts = (0, plan.horizon - 1)
horizon = active_ts[1] - active_ts[0] + 1
self._param_to_ll = {}
self.ll_start = active_ts[0]
for param in plan.params.values():
ll_param = LLParam(model, param, horizon, active_ts)
ll_param.create_grb_vars()
self._param_to_ll[param] = ll_param
model.update()
for ll_param in self._param_to_ll.values():
ll_param.batch_add_cnts()
def _add_obj_bexprs(self, obj_bexprs):
for bexpr in obj_bexprs:
self._prob.add_obj_expr(bexpr)
def _get_trajopt_obj(self, plan, active_ts=None):
if active_ts == None:
active_ts = (0, plan.horizon - 1)
start, end = active_ts
traj_objs = []
for param in plan.params.values():
if param not in self._param_to_ll:
continue
if param._type in ['Robot', 'Can']:
for attr_name in param.__dict__.iterkeys():
attr_type = param.get_attr_type(attr_name)
if issubclass(attr_type, Vector):
T = end - start + 1
K = attr_type.dim
attr_val = getattr(param, attr_name)
KT = K * T
v = -1 * np.ones((KT - K, 1))
d = np.vstack((np.ones((KT - K, 1)), np.zeros((K, 1))))
# [:,0] allows numpy to see v and d as one-dimensional so
# that numpy will create a diagonal matrix with v and d as a diagonal
P = np.diag(v[:, 0], K) + np.diag(d[:, 0])
Q = np.dot(np.transpose(P), P)
quad_expr = QuadExpr(Q, np.zeros((1, KT)),
np.zeros((1, 1)))
robot_ll = self._param_to_ll[param]
ll_attr_val = getattr(robot_ll, attr_name)
robot_ll_grb_vars = ll_attr_val.reshape((KT, 1),
order='F')
# init_val = attr_val[:, start:end+1].reshape((KT, 1), order='F')
# bexpr = BoundExpr(quad_expr, Variable(robot_ll_grb_vars, init_val))
bexpr = BoundExpr(quad_expr,
Variable(robot_ll_grb_vars))
traj_objs.append(bexpr)
return traj_objs
def _spawn_sco_var_for_pred(self, pred, t):
i = 0
x = np.empty(pred.x_dim, dtype=object)
v = np.empty(pred.x_dim)
for p in pred.attr_inds:
for attr, ind_arr in pred.attr_inds[p]:
n_vals = len(ind_arr)
ll_p = self._param_to_ll[p]
if p.is_symbol():
x[i:i + n_vals] = getattr(ll_p, attr)[ind_arr, 0]
v[i:i + n_vals] = getattr(p, attr)[ind_arr, 0]
else:
x[i:i + n_vals] = getattr(ll_p, attr)[ind_arr,
t - self.ll_start]
v[i:i + n_vals] = getattr(p, attr)[ind_arr, t]
i += n_vals
if pred.dynamic:
## include the variables from the next time step
for p in pred.attr_inds:
for attr, ind_arr in pred.attr_inds[p]:
n_vals = len(ind_arr)
ll_p = self._param_to_ll[p]
if p.is_symbol():
x[i:i + n_vals] = getattr(ll_p, attr)[ind_arr, 0]
v[i:i + n_vals] = getattr(p, attr)[ind_arr, 0]
else:
x[i:i + n_vals] = getattr(ll_p,
attr)[ind_arr,
t + 1 - self.ll_start]
v[i:i + n_vals] = getattr(p, attr)[ind_arr, t]
i += n_vals
assert i >= pred.x_dim
x = x.reshape((pred.x_dim, 1))
v = v.reshape((pred.x_dim, 1))
return Variable(x, v)
def _solve_opt_prob(self, plan, priority, callback=None, init=True,
active_ts=None, verbose=False):
robot = plan.params['baxter']
body = plan.env.GetRobot("baxter")
if callback is not None:
viewer = callback()
def draw(t):
viewer.draw_plan_ts(plan, t)
## active_ts is the inclusive timesteps to include
## in the optimization
if active_ts==None:
active_ts = (0, plan.horizon-1)
plan.save_free_attrs()
model = grb.Model()
model.params.OutputFlag = 0
self._prob = Prob(model, callback=callback)
# _free_attrs is paied attentioned in here
self._spawn_parameter_to_ll_mapping(model, plan, active_ts)
model.update()
self._bexpr_to_pred = {}
if priority == -2:
"""
Initialize an linear trajectory while enforceing the linear constraints in the intermediate step.
"""
obj_bexprs = self._get_trajopt_obj(plan, active_ts)
self._add_obj_bexprs(obj_bexprs)
self._add_first_and_last_timesteps_of_actions(plan,
priority=MAX_PRIORITY, active_ts=active_ts, verbose=verbose, add_nonlin=False)
tol = 1e-1
elif priority == -1:
"""
Solve the optimization problem while enforcing every constraints.
"""
obj_bexprs = self._get_trajopt_obj(plan, active_ts)
self._add_obj_bexprs(obj_bexprs)
self._add_first_and_last_timesteps_of_actions(plan,
priority=MAX_PRIORITY, active_ts=active_ts, verbose=verbose,
add_nonlin=True)
tol = 1e-1
elif priority == 0:
"""
When Optimization fails, resample new values for certain timesteps
of the trajectory and solver as initialization
"""
## this should only get called with a full plan for now
# assert active_ts == (0, plan.horizon-1)
failed_preds = plan.get_failed_preds()
# print "{} predicates fails, resampling process begin...\n \
# Checking {}".format(len(failed_preds), failed_preds[0])
## this is an objective that places
## a high value on matching the resampled values
obj_bexprs = []
rs_obj = self._resample(plan, failed_preds)
obj_bexprs.extend(rs_obj)
# _get_transfer_obj returns the expression saying the current trajectory should be close to it's previous trajectory.
# obj_bexprs.extend(self._get_trajopt_obj(plan, active_ts))
obj_bexprs.extend(self._get_transfer_obj(plan, self.transfer_norm))
self._add_obj_bexprs(obj_bexprs)
self._add_all_timesteps_of_actions(plan, priority=1,
add_nonlin=True, active_ts= active_ts, verbose=verbose)
tol = 1e-3
elif priority >= 1:
obj_bexprs = self._get_trajopt_obj(plan, active_ts)
self._add_obj_bexprs(obj_bexprs)
self._add_all_timesteps_of_actions(plan, priority=priority, add_nonlin=True,
active_ts=active_ts, verbose=verbose)
tol=1e-3
solv = Solver()
solv.initial_trust_region_size = self.initial_trust_region_size
solv.initial_penalty_coeff = self.init_penalty_coeff
solv.max_merit_coeff_increases = self.max_merit_coeff_increases
success = solv.solve(self._prob, method='penalty_sqp', tol=tol, verbose=True)
self._update_ll_params()
print "priority: {}".format(priority)
if priority == 0:
# During resampling phases, there must be changes added to sampling_trace
if len(plan.sampling_trace) > 0 and 'reward' not in plan.sampling_trace[-1]:
reward = 0
if len(plan.get_failed_preds()) == 0:
reward = len(plan.actions)
else:
failed_t = plan.get_failed_pred()[2]
for i in range(len(plan.actions)):
if failed_t > plan.actions[i].active_timesteps[1]:
reward += 1
plan.sampling_trace[-1]['reward'] = reward
##Restore free_attrs values
plan.restore_free_attrs()
return success
class RobotLLSolver(LLSolver):
def __init__(self, early_converge=False, transfer_norm='min-vel'):
self.transfer_coeff = 1e1
self.rs_coeff = 1e10
self.initial_trust_region_size = 1e-2
self.init_penalty_coeff = 1e1
# self.init_penalty_coeff = 1e5
self.max_merit_coeff_increases = 5
self._param_to_ll = {}
self.early_converge=early_converge
self.child_solver = None
self.solve_priorities = [2]
self.transfer_norm = transfer_norm
def _solve_helper(self, plan, callback, active_ts, verbose):
# certain constraints should be solved first
success = False
for priority in self.solve_priorities:
success = self._solve_opt_prob(plan, priority=priority,
callback=callback, active_ts=active_ts,
verbose=verbose)
# if not success:
# return success
# return True
return success
def solve(self, plan, callback=None, n_resamples=20, active_ts=None,
verbose=False, force_init=False):
success = False
if force_init or not plan.initialized:
## solve at priority -1 to get an initial value for the parameters
self._solve_opt_prob(plan, priority=-2, callback=callback,
active_ts=active_ts, verbose=verbose)
self._solve_opt_prob(plan, priority=-1, callback=callback,
active_ts=active_ts, verbose=verbose)
plan.initialized=True
success = self._solve_helper(plan, callback=callback,
active_ts=active_ts, verbose=verbose)
# self.saver = PlanSerializer()
# self.saver.write_plan_to_hdf5("temp_plan.hdf5", plan)
if success or len(plan.get_failed_preds()) == 0:
return True
# return success
for _ in range(n_resamples):
## refinement loop
## priority 0 resamples the first failed predicate in the plan
## and then solves a transfer optimization that only includes linear constraints
self._solve_opt_prob(plan, priority=0, callback=callback,
active_ts=active_ts, verbose=verbose)
success = self._solve_opt_prob(plan, priority=1,
callback=callback, active_ts=active_ts, verbose=verbose)
if success or len(plan.get_failed_preds()) == 0:
print "Optimization success after {} resampling.".format(_)
return _
return success
def _solve_opt_prob(self, plan, priority, callback=None, init=True,
active_ts=None, verbose=False):
robot = plan.params['baxter']
body = plan.env.GetRobot("baxter")
if callback is not None:
viewer = callback()
def draw(t):
viewer.draw_plan_ts(plan, t)
## active_ts is the inclusive timesteps to include
## in the optimization
if active_ts==None:
active_ts = (0, plan.horizon-1)
plan.save_free_attrs()
model = grb.Model()
model.params.OutputFlag = 0
self._prob = Prob(model, callback=callback)
# _free_attrs is paied attentioned in here
self._spawn_parameter_to_ll_mapping(model, plan, active_ts)
model.update()
self._bexpr_to_pred = {}
if priority == -2:
"""
Initialize an linear trajectory while enforceing the linear constraints in the intermediate step.
"""
obj_bexprs = self._get_trajopt_obj(plan, active_ts)
self._add_obj_bexprs(obj_bexprs)
self._add_first_and_last_timesteps_of_actions(plan,
priority=MAX_PRIORITY, active_ts=active_ts, verbose=verbose, add_nonlin=False)
tol = 1e-1
elif priority == -1:
"""
Solve the optimization problem while enforcing every constraints.
"""
obj_bexprs = self._get_trajopt_obj(plan, active_ts)
self._add_obj_bexprs(obj_bexprs)
self._add_first_and_last_timesteps_of_actions(plan,
priority=MAX_PRIORITY, active_ts=active_ts, verbose=verbose,
add_nonlin=True)
tol = 1e-1
elif priority == 0:
"""
When Optimization fails, resample new values for certain timesteps
of the trajectory and solver as initialization
"""
## this should only get called with a full plan for now
# assert active_ts == (0, plan.horizon-1)
failed_preds = plan.get_failed_preds()
# print "{} predicates fails, resampling process begin...\n \
# Checking {}".format(len(failed_preds), failed_preds[0])
## this is an objective that places
## a high value on matching the resampled values
obj_bexprs = []
rs_obj = self._resample(plan, failed_preds)
obj_bexprs.extend(rs_obj)
# _get_transfer_obj returns the expression saying the current trajectory should be close to it's previous trajectory.
# obj_bexprs.extend(self._get_trajopt_obj(plan, active_ts))
obj_bexprs.extend(self._get_transfer_obj(plan, self.transfer_norm))
self._add_obj_bexprs(obj_bexprs)
self._add_all_timesteps_of_actions(plan, priority=1,
add_nonlin=True, active_ts= active_ts, verbose=verbose)
tol = 1e-3
elif priority >= 1:
obj_bexprs = self._get_trajopt_obj(plan, active_ts)
self._add_obj_bexprs(obj_bexprs)
self._add_all_timesteps_of_actions(plan, priority=priority, add_nonlin=True,
active_ts=active_ts, verbose=verbose)
tol=1e-3
solv = Solver()
solv.initial_trust_region_size = self.initial_trust_region_size
solv.initial_penalty_coeff = self.init_penalty_coeff
solv.max_merit_coeff_increases = self.max_merit_coeff_increases
success = solv.solve(self._prob, method='penalty_sqp', tol=tol, verbose=True)
self._update_ll_params()
print "priority: {}".format(priority)
if priority == 0:
# During resampling phases, there must be changes added to sampling_trace
if len(plan.sampling_trace) > 0 and 'reward' not in plan.sampling_trace[-1]:
reward = 0
if len(plan.get_failed_preds()) == 0:
reward = len(plan.actions)
else:
failed_t = plan.get_failed_pred()[2]
for i in range(len(plan.actions)):
if failed_t > plan.actions[i].active_timesteps[1]:
reward += 1
plan.sampling_trace[-1]['reward'] = reward
##Restore free_attrs values
plan.restore_free_attrs()
return success
def _get_transfer_obj(self, plan, norm):
"""
This function returns the expression e(x) = P|x - cur|^2
Which says the optimized trajectory should be close to the
previous trajectory.
Where P is the KT x KT matrix, where Px is the difference of parameter's attributes' current value and parameter's next timestep value
"""
transfer_objs = []
if norm == 'min-vel':
for param in plan.params.values():
# if param._type in ['Robot', 'Can', 'EEPose']:
for attr_name in param.__dict__.iterkeys():
attr_type = param.get_attr_type(attr_name)
if issubclass(attr_type, Vector):
if param.is_symbol():
T = 1
else:
T = plan.horizon
K = attr_type.dim
attr_val = getattr(param, attr_name)
# pose = param.pose
assert (K, T) == attr_val.shape
KT = K*T
v = -1 * np.ones((KT - K, 1))
d = np.vstack((np.ones((KT - K, 1)), np.zeros((K, 1))))
# [:,0] allows numpy to see v and d as one-dimensional so
# that numpy will create a diagonal matrix with v and d as a diagonal
P = np.diag(v[:, 0], K) + np.diag(d[:, 0])
# P = np.eye(KT)
Q = np.dot(np.transpose(P), P)
cur_val = attr_val.reshape((KT, 1), order='F')
A = -2*cur_val.T.dot(Q)
b = cur_val.T.dot(Q.dot(cur_val))
# QuadExpr is 0.5*x^Tx + Ax + b
quad_expr = QuadExpr(2*self.transfer_coeff*Q,
self.transfer_coeff*A, self.transfer_coeff*b)
param_ll = self._param_to_ll[param]
ll_attr_val = getattr(param_ll, attr_name)
param_ll_grb_vars = ll_attr_val.reshape((KT, 1), order='F')
bexpr = BoundExpr(quad_expr, Variable(param_ll_grb_vars, cur_val))
transfer_objs.append(bexpr)
else:
raise NotImplemented
return transfer_objs
def _resample(self, plan, preds):
"""
This function first calls fail predicate's resample function,
then, uses the resampled value to create a square difference cost
function e(x) = |x - rs_val|^2 that will be minimized later.
rs_val is the resampled value
"""
val, attr_inds = None, None
for negated, pred, t in preds:
## returns a vector of new values and an
## attr_inds (OrderedDict) that gives the mapping
## to parameter attributes
val, attr_inds = pred.resample(negated, t, plan)
## if no resample defined for that pred, continue
if val is not None: break
if val is None:
return []
bexprs, i = [], 0
for p in attr_inds:
## get the ll_param for p and gurobi variables
ll_p = self._param_to_ll[p]
n_vals = 0
grb_vars = []
for attr, ind_arr, t in attr_inds[p]:
n_vals += len(ind_arr)
grb_vars.extend(
list(getattr(ll_p, attr)[ind_arr, t].flatten()))
for j, grb_var in enumerate(grb_vars):
## create an objective saying stay close to the resampled value
## e(x) = (x - val[i+j])**2
## e(x) = x^2 - 2*val[i+j]*x + val[i+j]^2
Q = np.eye(1)
A = -2*val[i+j]*np.ones((1, 1))
b = np.ones((1, 1))*np.power(val[i+j], 2)
# QuadExpr is 0.5*x^Tx + Ax + b
quad_expr = QuadExpr(2*Q*self.rs_coeff, A*self.rs_coeff, b*self.rs_coeff)
v_arr = np.array([grb_var]).reshape((1, 1), order='F')
init_val = np.ones((1, 1))*val[i+j]
bexpr = BoundExpr(quad_expr,
Variable(v_arr, val[i+j].reshape((1, 1))))
bexprs.append(bexpr)
i += n_vals
return bexprs
def _add_pred_dict(self, pred_dict, effective_timesteps, add_nonlin=True,
priority=MAX_PRIORITY, verbose=False):
"""
This function creates constraints for the predicate and added to
Prob class in sco.
"""
## for debugging
ignore_preds = []
priority = np.maximum(priority, 0)
if not pred_dict['hl_info'] == "hl_state":
start, end = pred_dict['active_timesteps']
active_range = range(start, end+1)
if verbose:
print "pred being added: ", pred_dict
print active_range, effective_timesteps
negated = pred_dict['negated']
pred = pred_dict['pred']
if pred.get_type() in ignore_preds:
return
if pred.priority > priority: return
assert isinstance(pred, common_predicates.ExprPredicate)
expr = pred.get_expr(negated)
for t in effective_timesteps:
if t in active_range:
if expr is not None:
if add_nonlin or isinstance(expr.expr, AffExpr):
if verbose:
print "expr being added at time ", t
var = self._spawn_sco_var_for_pred(pred, t)
bexpr = BoundExpr(expr, var)
# TODO: REMOVE line below, for tracing back predicate for debugging.
bexpr.pred = pred
self._bexpr_to_pred[bexpr] = (negated, pred, t)
groups = ['all']
if self.early_converge:
## this will check for convergence per parameter
## this is good if e.g., a single trajectory quickly
## gets stuck
groups.extend([param.name for param in pred.params])
self._prob.add_cnt_expr(bexpr, groups)
def _add_first_and_last_timesteps_of_actions(self, plan, priority = MAX_PRIORITY,
add_nonlin=False, active_ts=None, verbose=False):
"""
This function adds all linear predicates and first and last timestep
non-linear predicates from actions that are active within the range of active_ts.
"""
if active_ts==None:
active_ts = (0, plan.horizon-1)
for action in plan.actions:
action_start, action_end = action.active_timesteps
## only add an action
if action_start >= active_ts[1] and action_start > active_ts[0]: continue
if action_end < active_ts[0]: continue
for pred_dict in action.preds:
if action_start >= active_ts[0]:
self._add_pred_dict(pred_dict, [action_start],
priority=priority, add_nonlin=add_nonlin, verbose=verbose)
if action_end <= active_ts[1]:
self._add_pred_dict(pred_dict, [action_end],
priority=priority, add_nonlin=add_nonlin, verbose=verbose)
## add all of the linear ineqs
timesteps = range(max(action_start+1, active_ts[0]),
min(action_end, active_ts[1]))
for pred_dict in action.preds:
self._add_pred_dict(pred_dict, timesteps, add_nonlin=False,
priority=priority, verbose=verbose)
def _add_all_timesteps_of_actions(self, plan, priority=MAX_PRIORITY,
add_nonlin=True, active_ts=None, verbose=False):
"""
This function adds both linear and non-linear predicates from
actions that are active within the range of active_ts.
"""
if active_ts==None:
active_ts = (0, plan.horizon-1)
for action in plan.actions:
action_start, action_end = action.active_timesteps
if action_start >= active_ts[1] and action_start > active_ts[0]: continue
if action_end < active_ts[0]: continue
timesteps = range(max(action_start, active_ts[0]),
min(action_end, active_ts[1])+1)
for pred_dict in action.preds:
self._add_pred_dict(pred_dict, timesteps, priority=priority,
add_nonlin=add_nonlin, verbose=verbose)
def _update_ll_params(self):
"""
update plan's parameters from low level grb_vars.
expected to be called after each optimization.
"""
for ll_param in self._param_to_ll.values():
ll_param.update_param()
if self.child_solver:
self.child_solver._update_ll_params()
def _spawn_parameter_to_ll_mapping(self, model, plan, active_ts=None):
"""
This function creates low level parameters for each parameter in the plan,
initialized he corresponding grb_vars for each attributes in each timestep,
update the grb models
adds in equality constraints,
construct a dictionary as param-to-ll_param mapping.
"""
if active_ts == None:
active_ts=(0, plan.horizon-1)
horizon = active_ts[1] - active_ts[0] + 1
self._param_to_ll = {}
self.ll_start = active_ts[0]
for param in plan.params.values():
ll_param = LLParam(model, param, horizon, active_ts)
ll_param.create_grb_vars()
self._param_to_ll[param] = ll_param
model.update()
for ll_param in self._param_to_ll.values():
ll_param.batch_add_cnts()
def _add_obj_bexprs(self, obj_bexprs):
"""
This function adds objective bounded expressions to the Prob class
in sco.
"""
for bexpr in obj_bexprs:
self._prob.add_obj_expr(bexpr)
def _get_trajopt_obj(self, plan, active_ts=None):
"""
This function selects parameter of type Robot and Can and returns
the expression e(x) = |Px|^2
Which optimize trajectory so that robot and can's attributes in
current timestep is close to that of next timestep.
forming a straight line between each end points.
Where P is the KT x KT matrix, where Px is the difference of
value in current timestep compare to next timestep
"""
if active_ts == None:
active_ts = (0, plan.horizon-1)
start, end = active_ts
traj_objs = []
for param in plan.params.values():
if param not in self._param_to_ll:
continue
if param._type in ['Robot', 'Can']:
for attr_name in param.__dict__.iterkeys():
attr_type = param.get_attr_type(attr_name)
if issubclass(attr_type, Vector):
T = end - start + 1
K = attr_type.dim
attr_val = getattr(param, attr_name)
KT = K*T
v = -1 * np.ones((KT - K, 1))
d = np.vstack((np.ones((KT - K, 1)), np.zeros((K, 1))))
# [:,0] allows numpy to see v and d as one-dimensional so
# that numpy will create a diagonal matrix with v and d as a diagonal
P = np.diag(v[:, 0], K) + np.diag(d[:, 0])
Q = np.dot(np.transpose(P), P)
Q *= TRAJOPT_COEFF
quad_expr = None
if attr_name == 'pose' and param._type == 'Robot':
quad_expr = QuadExpr(BASE_MOVE_COEFF*Q,
np.zeros((1,KT)),
np.zeros((1,1)))
else:
quad_expr = QuadExpr(Q, np.zeros((1,KT)), np.zeros((1,1)))
param_ll = self._param_to_ll[param]
ll_attr_val = getattr(param_ll, attr_name)
param_ll_grb_vars = ll_attr_val.reshape((KT, 1), order='F')
attr_val = getattr(param, attr_name)
init_val = attr_val[:, start:end+1].reshape((KT, 1), order='F')
bexpr = BoundExpr(quad_expr, Variable(param_ll_grb_vars, init_val))
# bexpr = BoundExpr(quad_expr, Variable(param_ll_grb_vars))
traj_objs.append(bexpr)
return traj_objs