def test_model(sampler, args, gripper=None):
"""
Maximize the GP mean function to get the best policy.
:param sampler: A GP fit to the current BusyBox.
:return: Regret.
"""
# Optimize the GP to get the best policy.
ucb = False
stop_x, stop_policy, start_x = sampler.optim.optimize_gp(ucb)
# Execute the policy and observe the true motion.
debug = False
viz = False
if gripper is None:
_, gripper = setup_env(sampler.bb, viz, debug)
else:
gripper.reset(sampler.mech)
pose_handle_base_world = sampler.mech.get_pose_handle_base_world()
traj = stop_policy.generate_trajectory(pose_handle_base_world, debug=debug)
cmotion, motion, _ = gripper.execute_trajectory(traj, sampler.mech, stop_policy.type, debug=debug)
# Calculate the regret.
max_d = sampler.mech.get_max_net_motion()
regret = (max_d - motion)/max_d
return regret, start_x, stop_x, stop_policy.type
def main(args):
busybox_data = get_bb_dataset(args.bb_fname, args.N, args.mech_types, 1,
args.urdf_num)
all_steps = []
for ix, bb_results in enumerate(busybox_data):
if args.type == 'gpucb':
single_dataset, _, steps = create_single_bb_gpucb_dataset(
bb_results[0],
'',
args.plot,
args,
ix,
success_regret=SUCCESS_REGRET)
all_steps.append(steps)
print('steps', steps)
elif args.type == 'random':
bb = BusyBox.bb_from_result(bb_results[0])
image_data, gripper = setup_env(bb, args.viz, args.debug)
regret = float("inf")
steps = 0
while regret > SUCCESS_REGRET:
mech = bb._mechanisms[0]
# generate either a random or model-based policy and goal configuration
policy = policies.generate_policy(mech)
pose_handle_world_init = mech.get_handle_pose()
# calculate trajectory
pose_handle_base_world = mech.get_pose_handle_base_world()
traj = policy.generate_trajectory(pose_handle_base_world,
args.debug,
color=[0, 0, 1])
# execute trajectory
cumu_motion, net_motion, pose_handle_world_final = \
gripper.execute_trajectory(traj, mech, policy.type, args.debug)
# calc regret
max_dist = mech.get_max_net_motion()
regret = (max_dist - net_motion) / max_dist
steps += 1
# reset
gripper.reset(mech)
p.disconnect()
all_steps.append(steps)
print('steps', steps)
# Save the dataset.
util.write_to_file(args.fname, all_steps)
def replay_result(result):
from gen.generator_busybox import BusyBox
from utils.setup_pybullet import setup_env
from actions.policies import get_policy_from_tuple
bb = BusyBox.bb_from_result(result)
image_data, gripper = setup_env(bb, True, True)
mech = bb._mechanisms[0]
policy = get_policy_from_tuple(result.policy_params)
pose_handle_base_world = mech.get_pose_handle_base_world()
traj = policy.generate_trajectory(pose_handle_base_world, True)
cumu_motion, net_motion, pose_handle_world_final = gripper.execute_trajectory(
traj, mech, policy.type, True)
import pdb
pdb.set_trace()
p.disconnect()
def get_true_ys(X_pred, mech, policy_params):
Y_pred = np.zeros((X_pred.shape[0]))
viz = False
debug = False
for i, x in enumerate(X_pred):
policy = policies.get_policy_from_x(mech, x, policy_params)
width, height = 0.6, 0.6
bb = BusyBox.get_busybox(width, height, [mech])
_, gripper = setup_env(bb, viz, debug)
# calculate trajectory
pose_handle_base_world = mech.get_pose_handle_base_world()
traj = policy.generate_trajectory(pose_handle_base_world, debug)
# execute trajectory
_, net_motion, _ = gripper.execute_trajectory(traj, mech, policy.type, False)
Y_pred[i] = net_motion
gripper.reset(mech)
return Y_pred
def generate_dataset(args, git_hash):
bb_dataset = get_bb_dataset(args.bb_fname, args.n_bbs, args.mech_types, args.max_mech, args.urdf_num)
if args.n_samples == 0:
return bb_dataset
results = []
for (i, bb_results) in enumerate(bb_dataset):
bb = BusyBox.bb_from_result(bb_results[0])
image_data, gripper = setup_env(bb, args.viz, args.debug)
bb_results = []
for j in range(args.n_samples):
sys.stdout.write("\rProcessing sample %i/%i for busybox %i/%i" % (j+1, args.n_samples, i+1, args.n_bbs))
for mech in bb._mechanisms:
# generate either a random or model-based policy and goal configuration
policy = policies.generate_policy(mech)
pose_handle_world_init = mech.get_handle_pose()
# calculate trajectory
pose_handle_base_world = mech.get_pose_handle_base_world()
traj = policy.generate_trajectory(pose_handle_base_world, args.debug, color=[0, 0, 1])
# execute trajectory
cumu_motion, net_motion, pose_handle_world_final = \
gripper.execute_trajectory(traj, mech, policy.type, args.debug)
# save result data
policy_params = policy.get_policy_tuple()
mechanism_params = mech.get_mechanism_tuple()
result = util.Result(policy_params, mechanism_params, net_motion, \
cumu_motion, pose_handle_world_init, pose_handle_world_final, \
image_data, git_hash)
bb_results.append(result)
gripper.reset(mech)
results.append(bb_results)
p.disconnect()
print()
return results
def get_bb_dataset(bb_fname, n_bbs, mech_types, max_mech, urdf_num):
# Create a dataset of busyboxes.
if bb_fname == '' or bb_fname is None:
print('Creating Busyboxes.')
mech_classes = []
for mech_type in mech_types:
if mech_type == 'slider': mech_classes.append(Slider)
if mech_type == 'door': mech_classes.append(Door)
bb_dataset = []
for _ in range(n_bbs):
# TODO: i think there is a bug here...
bb = BusyBox.generate_random_busybox(max_mech=max_mech,
mech_types=mech_classes,
urdf_tag=urdf_num)
mechanism_params = bb._mechanisms[0].get_mechanism_tuple()
image_data, gripper = setup_env(bb, False, False)
bb_dataset.append([util.Result(None, mechanism_params, None, None, None,
None, image_data, None)])
print('BusyBoxes created.')
else:
# Load in a file with predetermined BusyBoxes.
bb_dataset = read_from_file(bb_fname)[:n_bbs]
return bb_dataset
def create_single_bb_gpucb_dataset(bb_result, nn_fname, plot, args, bb_i,
n_interactions=None, plot_dir_prefix='',
ret_regret=False, success_regret=None):
use_cuda = False
dataset = []
viz = False
debug = False
# interact with BB
bb = BusyBox.bb_from_result(bb_result, urdf_num=args.urdf_num)
mech = bb._mechanisms[0]
image_data, gripper = setup_env(bb, viz, debug)
pose_handle_base_world = mech.get_pose_handle_base_world()
sampler = UCB_Interaction(bb, image_data, plot, args, nn_fname=nn_fname)
for ix in itertools.count():
gripper.reset(mech)
if args.debug:
sys.stdout.write('\rProcessing sample %i' % ix)
# if we are done sampling n_interactions OR we need to get regret after
# each interaction
if ((not n_interactions is None) and ix==n_interactions) or \
(not success_regret is None):
regret, start_x, stop_x, policy_type = test_model(sampler, args, gripper=gripper)
gripper.reset(mech)
#print('Current regret', regret)
opt_points = (policy_type, [(start_x, 'g'), (stop_x, 'r')])
sample_points = {'Prismatic': [(sample, 'k') for sample in sampler.xs['Prismatic']],
'Revolute': [(sample, 'k') for sample in sampler.xs['Revolute']]}
# if done sampling n_interactions
if (not n_interactions is None) and ix==n_interactions:
if plot:
viz_circles(util.GP_PLOT,
image_data,
mech,
BETA,
sample_points=sample_points,
opt_points=opt_points,
gps=sampler.gps,
nn=sampler.nn,
bb_i=bb_i,
plot_dir_prefix=plot_dir_prefix)
plt.show()
input('Enter to close')
plt.close()
if ret_regret:
return dataset, sampler.gps, regret
else:
return dataset, sampler.gps
# if got successful interaction or timeout
elif (not success_regret is None) and \
((regret < success_regret) or (ix >= 100)):
if ix >= 100:
print('timeout interactions')
elif regret < success_regret:
print('succcessful interaction!')
if plot:
viz_circles(util.GP_PLOT,
image_data,
mech,
BETA,
sample_points=sample_points,
opt_points=opt_points,
gps=sampler.gps,
nn=sampler.nn,
bb_i=bb_i,
plot_dir_prefix=plot_dir_prefix)
return dataset, sampler.gps, ix
# sample a policy
# image_data, gripper = setup_env(bb, False, debug)
gripper.reset(mech)
x, policy = sampler.sample()
# execute
traj = policy.generate_trajectory(pose_handle_base_world, debug=debug)
c_motion, motion, handle_pose_final = gripper.execute_trajectory(traj, mech, policy.type, False)
result = util.Result(policy.get_policy_tuple(), mech.get_mechanism_tuple(),
motion, c_motion, handle_pose_final, handle_pose_final,
image_data, None)
dataset.append(result)
# update GP
sampler.update(result, x)