class Robot(object):
mass = None
def __init__(self, env, robot_name):
env.GetPhysicsEngine().SetGravity(array([0, 0, -9.81]))
rave = env.GetRobot(robot_name)
q_min, q_max = rave.GetDOFLimits()
nb_dofs = rave.GetDOF()
rave.SetDOFVelocityLimits(1000 * rave.GetDOFVelocityLimits())
rave.SetDOFVelocities([0] * nb_dofs)
self.active_dofs = None
self.env = env
if self.mass is None: # may not be True for children classes
self.mass = sum([link.GetMass() for link in rave.GetLinks()])
self.nb_dofs = nb_dofs
self.nb_active_dofs = rave.GetDOF()
self.q_max_full = q_max
self.q_min_full = q_min
self.rave = rave
#
# Kinematics
#
def set_active_dofs(self, active_dofs):
self.active_dofs = active_dofs
self.nb_active_dofs = len(active_dofs)
self.rave.SetActiveDOFs(active_dofs)
def get_dof_values(self, dof_indices=None):
if dof_indices is not None:
return self.rave.GetDOFValues(dof_indices)
elif self.active_dofs:
return self.rave.GetActiveDOFValues()
return self.rave.GetDOFValues()
def set_dof_values(self, q, dof_indices=None):
if dof_indices is not None:
return self.rave.SetDOFValues(q, dof_indices)
elif self.active_dofs and len(q) == self.nb_active_dofs:
return self.rave.SetDOFValues(q, self.active_dofs)
assert len(q) == self.nb_dofs
return self.rave.SetDOFValues(q)
@property
def q(self):
return self.get_dof_values()
@property
def q_full(self):
return self.rave.GetDOFValues()
def scale_dof_limits(self, scale=1.):
q_avg = .5 * (self.q_max_full + self.q_min_full)
q_dev = .5 * (self.q_max_full - self.q_min_full)
self.q_max_full = (q_avg + scale * q_dev)
self.q_min_full = (q_avg - scale * q_dev)
@property
def q_max(self):
if not self.active_dofs:
return self.q_max_full
return self.q_max_full[self.active_dofs]
@property
def q_min(self):
if not self.active_dofs:
return self.q_min_full
return self.q_min_full[self.active_dofs]
def get_dof_velocities(self, dof_indices=None):
if dof_indices is not None:
return self.rave.GetDOFVelocities(dof_indices)
elif self.active_dofs:
return self.rave.GetActiveDOFVelocities()
return self.rave.GetDOFVelocities()
def set_dof_velocities(self, qd, dof_indices=None):
check_dof_limits = 0 # CLA_Nothing
if dof_indices is not None:
return self.rave.SetDOFVelocities(qd, check_dof_limits, dof_indices)
elif self.active_dofs and len(qd) == self.nb_active_dofs:
return self.rave.SetDOFVelocities(qd, check_dof_limits,
self.active_dofs)
assert len(qd) == self.nb_dofs
return self.rave.SetDOFVelocities(qd)
@property
def qd(self):
return self.get_dof_velocities()
@property
def qd_full(self):
return self.rave.GetDOFVelocities()
#
# Inverse Kinematics
#
def init_ik(self, dt, qd_lim, K_doflim, doflim_scale=0.95):
self.ik = DiffIKSolver(self, dt, qd_lim, K_doflim, doflim_scale)
def add_com_objective(self, target, gain, weight):
def error(q, qd):
return target.p - self.compute_com(q)
self.ik.add_objective(error, self.compute_com_jacobian, gain, weight)
def add_link_objective(self, link, target, gain, weight):
def error(q, qd):
cur_pose = self.compute_link_pose(link, q, self.active_dofs)
return target.pose - cur_pose
def jacobian(q):
return self.compute_link_pose_jacobian(link, q)
self.ik.add_objective(error, jacobian, gain, weight)
def add_link_pos_objective(self, link, target, gain, weight):
def error(q, qd):
return target.p - link.p
def jacobian(q):
return self.compute_link_translation_jacobian(link, q)
self.ik.add_objective(error, jacobian, gain, weight)
def add_link_objective_2(self, link, target, alpha, gain, weight):
"""
This one is quite specific to my needs. Here, alpha is a callable
returning a float between zero and one.
"""
def error(q, qd):
cur_pose = self.compute_link_pose(link, q, self.active_dofs)
return alpha() * (target.pose - cur_pose)
def jacobian(q):
return self.compute_link_pose_jacobian(link, q)
self.ik.add_objective(error, jacobian, gain, weight)
def add_constant_cam_objective(self, weight):
def error(q, qd):
J = self.compute_cam_jacobian(q)
# Ld_G = J d(qd) / dt + qd * H * qd, regulated to 0
if False:
# i.e., J qd_new = J qd_prev - dt * qd_prev * H * qd_prev
H = self.compute_cam_hessian(q)
return dot(J, qd) - self.ik.dt * dot(qd, dot(H, qd))
else:
# neglecting the hessian term, this becomes
return dot(J, qd)
self.ik.add_objective(error, self.compute_cam_jacobian, 1., weight)
def add_zero_cam_objective(self, weight):
def error(q, qd):
return zeros((3,))
self.ik.add_objective(error, self.compute_cam_jacobian, 0., weight)
def add_posture_objective(self, q_ref, gain, weight):
if len(q_ref) == self.nb_dofs:
q_ref = q_ref[self.active_dofs]
assert len(q_ref) == self.nb_active_dofs
def error(q, qd):
return (q_ref - q)
self.ik.add_objective(error, self.ik.identity, gain, weight)
def add_dof_objective(self, dof_id, dof_ref, gain, weight):
active_dof_id = self.active_dofs.index(dof_id)
J = zeros(self.nb_active_dofs)
J[active_dof_id] = 1.
def error(q, qd):
return (dof_ref - q[active_dof_id])
def jacobian(q):
return J
self.ik.add_objective(error, jacobian, gain, weight)
def add_velocity_regularization(self, weight):
def error(q, qd):
return qd
self.ik.add_objective(error, self.ik.identity, 1., weight)
def add_zero_velocity_objective(self, gain, weight):
def error(q, qd):
return -qd
self.ik.add_objective(error, self.ik.identity, gain, weight)
def step_tracker(self):
qd = self.ik.compute_velocity(self.q, self.qd)
self.set_dof_values(self.q + qd * self.ik.dt)
self.set_dof_velocities(qd)
def solve_ik(self, max_it=100, conv_tol=1e-5, debug=True):
cur_obj = 1000.
q = self.q
qd = zeros(len(self.q))
if debug:
print "solve_ik(max_it=%d, conv_tol=%e)" % (max_it, conv_tol)
for itnum in xrange(max_it):
prev_obj = cur_obj
cur_obj = self.ik.compute_objective(q, qd)
if debug:
print "%2d: %.3f (%+.2e)" % (itnum, cur_obj, cur_obj - prev_obj)
if abs(cur_obj - prev_obj) < conv_tol:
break
qd = self.ik.compute_velocity(q, qd)
q = minimum(maximum(self.q_min, q + qd * self.ik.dt), self.q_max)
self.set_dof_values(q)
return itnum, cur_obj
#
# Visualization
#
def play_trajectory(self, traj, callback=None, dt=3e-2, dof_indices=None):
trange = list(arange(0, traj.T, dt))
for t in trange:
q = traj.q(t)
qd = traj.qd(t)
qdd = traj.qdd(t)
self.set_dof_values(q, dof_indices)
if callback:
callback(t, q, qd, qdd)
time.sleep(dt)
def record_trajectory(self, traj, fname='output.mpg', codec=13,
framerate=24, width=800, height=600, dt=3e-2):
viewer = self.env.GetViewer()
recorder = RaveCreateModule(self.env, 'viewerrecorder')
self.env.AddModule(recorder, '')
self.set_dof_values(traj.q(0))
recorder.SendCommand('Start %d %d %d codec %d timing '
'simtime filename %s\n'
'viewer %s' % (width, height, framerate, codec,
fname, viewer.GetName()))
time.sleep(1.)
self.play_trajectory(traj, dt=dt)
time.sleep(1.)
recorder.SendCommand('Stop')
self.env.Remove(recorder)
def set_color(self, r, g, b):
for link in self.rave.GetLinks():
for geom in link.GetGeometries():
geom.SetAmbientColor([r, g, b])
geom.SetDiffuseColor([r, g, b])
def set_transparency(self, transparency):
for link in self.rave.GetLinks():
for geom in link.GetGeometries():
geom.SetTransparency(transparency)
#
# Geometric properties (position level)
#
def compute_com(self, q, dof_indices=None):
total = zeros(3)
with self.rave:
self.set_dof_values(q, dof_indices)
for link in self.rave.GetLinks():
m = link.GetMass()
c = link.GetGlobalCOM()
total += m * c
return total / self.mass
@property
def com(self):
return self.compute_com(self.q)
def compute_inertia_matrix(self, q, dof_indices=None, external_torque=None):
M = zeros((self.nb_dofs, self.nb_dofs))
self.set_dof_values(q, dof_indices)
for (i, e_i) in enumerate(eye(self.nb_dofs)):
tm, _, _ = self.rave.ComputeInverseDynamics(
e_i, external_torque, returncomponents=True)
M[:, i] = tm
return M
def compute_link_pos(self, link, q, link_coord=None, dof_indices=None):
with self.rave:
self.set_dof_values(q, dof_indices)
T = link.T
if link_coord is None:
return T[:3, 3]
return dot(T, hstack([link_coord, 1]))[:3]
def compute_link_pose(self, link, q=None, dof_indices=None):
if q is None:
return link.pose
with self.rave:
self.set_dof_values(q, dof_indices)
self.rave.SetDOFValues(q, dof_indices)
return link.pose # first coefficient will be positive
#
# Kinematic properties (velocity level)
#
def compute_angular_momentum(self, q, qd, p, dof_indices=None):
"""Compute the angular momentum with respect to point p.
q -- joint angle values
qd -- joint-angle velocities
p -- application point, either a fixed point or the instantaneous COM,
in world coordinates
"""
momentum = zeros(3)
with self.rave:
self.set_dof_values(q, dof_indices)
self.set_dof_velocities(qd, dof_indices)
for link in self.rave.GetLinks():
T = link.GetTransform()
R, r = T[0:3, 0:3], T[0:3, 3]
c_local = link.GetLocalCOM() # in local RF
c = r + dot(R, c_local)
v = link.GetVelocity()
rd, omega = v[:3], v[3:]
cd = rd + cross(omega, dot(R, c_local))
m = link.GetMass()
I = link.GetLocalInertia() # in local RF
momentum += cross(c - p, m * cd) \
+ dot(R, dot(I, dot(R.T, omega)))
return momentum
def compute_cam(self, q, qd, dof_indices=None):
"""
Compute the centroidal angular momentum (CAM), that is to say, the
angular momentum with respect to the COM.
"""
p_G = self.compute_com(q, dof_indices)
return self.compute_angular_momentum(q, qd, p_G, dof_indices)
@property
def cam(self):
return self.compute_cam(self.q, self.qd)
def compute_com_velocity(self, q, qd, dof_indices=None):
total = zeros(3)
with self.rave:
self.set_dof_values(q, dof_indices)
self.set_dof_velocities(qd, dof_indices)
for link in self.rave.GetLinks():
m = link.GetMass()
R = link.GetTransform()[0:3, 0:3]
c_local = link.GetLocalCOM()
v = link.GetVelocity()
rd, omega = v[:3], v[3:]
cd = rd + cross(omega, dot(R, c_local))
total += m * cd
return total / self.mass
@property
def comd(self):
return self.compute_com_velocity(self.q, self.qd)
#
# Dynamic properties (acceleration level)
#
def compute_cam_rate(self, q, qd, qdd):
J = self.compute_cam_jacobian(q)
H = self.compute_cam_hessian(q)
return dot(J, qdd) + dot(qd, dot(H, qd))
def compute_com_acceleration(self, q, qd, qdd):
J = self.compute_com_jacobian(q)
H = self.compute_com_hessian(q)
return dot(J, qdd) + dot(qd, dot(H, qdd))
def compute_zmp(self, q, qd, qdd, dof_indices=None):
global pb_times, total_times, cum_ratio, avg_ratio
g = array([0, 0, -9.81])
f0 = self.mass * g[2]
tau0 = zeros(3)
with self.rave:
self.set_dof_values(q, dof_indices)
self.set_dof_velocities(qd, dof_indices)
link_velocities = self.rave.GetLinkVelocities()
link_accelerations = self.rave.GetLinkAccelerations(qdd)
for link in self.rave.GetLinks():
mi = link.GetMass()
ci = link.GetGlobalCOM()
I_ci = link.GetLocalInertia()
Ri = link.GetTransform()[0:3, 0:3]
ri = dot(Ri, link.GetLocalCOM())
# linvel = link_velocities[link.GetIndex()][:3]
angvel = link_velocities[link.GetIndex()][3:]
linacc = link_accelerations[link.GetIndex()][:3]
angacc = link_accelerations[link.GetIndex()][3:]
# ci_dot = linvel + cross(angvel, ri)
ci_ddot = linacc \
+ cross(angvel, cross(angvel, ri)) \
+ cross(angacc, ri)
angmmt = dot(I_ci, angacc) - cross(dot(I_ci, angvel), angvel)
f0 -= mi * ci_ddot[2]
tau0 += mi * cross(ci, g - ci_ddot) - dot(Ri, angmmt)
return cross(array([0, 0, 1]), tau0) * 1. / f0
#
# Jacobians
#
def compute_am_jacobian(self, q, p, dof_indices=None):
"""Compute a matrix J(p) such that the angular momentum with respect to
the point p is
L_p(q, qd) = dot(J(q), qd).
q -- joint angle values
qd -- joint-angle velocities
p -- application point, either a fixed point or the instantaneous COM,
in world coordinates
"""
J = zeros((3, self.nb_dofs))
with self.rave:
self.set_dof_values(q, dof_indices)
for link in self.rave.GetLinks():
m = link.GetMass()
i = link.GetIndex()
c = link.GetGlobalCOM()
R = link.GetTransform()[0:3, 0:3]
I = dot(R, dot(link.GetLocalInertia(), R.T))
J_trans = self.rave.ComputeJacobianTranslation(i, c)
J_rot = self.rave.ComputeJacobianAxisAngle(i)
J += dot(crossmat(c - p), m * J_trans) + dot(I, J_rot)
if dof_indices is not None:
return J[:, dof_indices]
elif self.active_dofs and len(q) == self.nb_active_dofs:
return J[:, self.active_dofs]
return J
def compute_cam_jacobian(self, q, dof_indices=None):
"""Compute a matrix J(p) such that the angular momentum with respect to
the center of mass G is
L_G(q, qd) = dot(J(q), qd).
q -- joint angle values
qd -- joint-angle velocities
"""
p_G = self.compute_com(q, dof_indices)
return self.compute_am_jacobian(q, p_G, dof_indices)
def compute_com_jacobian(self, q, dof_indices=None):
Jcom = zeros((3, self.nb_dofs))
with self.rave:
self.set_dof_values(q, dof_indices)
for link in self.rave.GetLinks():
index = link.GetIndex()
com = link.GetGlobalCOM()
m = link.GetMass()
J = self.rave.ComputeJacobianTranslation(index, com)
Jcom += m * J
J = Jcom / self.mass
if dof_indices is not None:
return J[:, dof_indices]
elif self.active_dofs and len(q) == self.nb_active_dofs:
return J[:, self.active_dofs]
return J
def compute_link_jacobian(self, link, q, dof_indices=None):
with self.rave:
self.set_dof_values(q, dof_indices)
J_trans = self.rave.ComputeJacobianTranslation(link.index, link.p)
J_rot = self.rave.ComputeJacobianAxisAngle(link.index)
J = vstack([J_rot, J_trans])
if dof_indices is not None:
return J[:, dof_indices]
elif self.active_dofs and len(q) == self.nb_active_dofs:
return J[:, self.active_dofs]
return J
def compute_link_pose_jacobian(self, link, q, dof_indices=None):
with self.rave:
self.set_dof_values(q, dof_indices)
J_trans = self.rave.CalculateJacobian(link.index, link.p)
or_quat = link.rave.GetTransformPose()[:4] # don't use link.pose
J_quat = self.rave.CalculateRotationJacobian(link.index, or_quat)
if or_quat[0] < 0: # we enforce positive first coefficients
J_quat *= -1.
J = vstack([J_quat, J_trans])
if dof_indices is not None:
return J[:, dof_indices]
elif self.active_dofs and len(q) == self.nb_active_dofs:
return J[:, self.active_dofs]
return J
def compute_link_translation_jacobian(self, link, q, link_coord=None,
dof_indices=None):
with self.rave:
self.set_dof_values(q, dof_indices)
p = self.compute_link_pos(link, q, link_coord, dof_indices)
J = self.rave.ComputeJacobianTranslation(link.index, p)
if dof_indices is not None:
return J[:, dof_indices]
elif self.active_dofs and len(q) == self.nb_active_dofs:
return J[:, self.active_dofs]
return J
#
# Hessians
#
def compute_am_hessian(self, q, p, dof_indices=None):
"""Returns a matrix H(q) such that the rate of change of the angular
momentum with respect to point p is
Ld_p(q, qd) = dot(J(q), qdd) + dot(qd.T, dot(H(q), qd)),
where J(q) is the result of self.compute_am_jacobian(q, p).
q -- joint angle values
qd -- joint-angle velocities
p -- application point, either a fixed point or the instantaneous COM,
in world coordinates
"""
def crosstens(M):
assert M.shape[0] == 3
Z = zeros(M.shape[1])
T = array([[Z, -M[2, :], M[1, :]],
[M[2, :], Z, -M[0, :]],
[-M[1, :], M[0, :], Z]])
return T.transpose([2, 0, 1]) # T.shape == (M.shape[1], 3, 3)
def middot(M, T):
"""Dot product of a matrix with the mid-coordinate of a 3D tensor.
M -- matrix with shape (n, m)
T -- tensor with shape (a, m, b)
Outputs a tensor of shape (a, n, b).
"""
return tensordot(M, T, axes=(1, 1)).transpose([1, 0, 2])
H = zeros((self.nb_dofs, 3, self.nb_dofs))
with self.rave:
self.set_dof_values(q)
for link in self.rave.GetLinks():
m = link.GetMass()
i = link.GetIndex()
c = link.GetGlobalCOM()
R = link.GetTransform()[0:3, 0:3]
# J_trans = self.rave.ComputeJacobianTranslation(i, c)
J_rot = self.rave.ComputeJacobianAxisAngle(i)
H_trans = self.rave.ComputeHessianTranslation(i, c)
H_rot = self.rave.ComputeHessianAxisAngle(i)
I = dot(R, dot(link.GetLocalInertia(), R.T))
H += middot(crossmat(c - p), m * H_trans) \
+ middot(I, H_rot) \
- dot(crosstens(dot(I, J_rot)), J_rot)
if dof_indices:
return ((H[dof_indices, :, :])[:, :, dof_indices])
elif self.active_dofs and len(q) == self.nb_active_dofs:
return ((H[self.active_dofs, :, :])[:, :, self.active_dofs])
return H
def compute_cam_hessian(self, q, dof_indices=None):
"""Returns a matrix H(q) such that the rate of change of the angular
momentum with respect to the center of mass G is
Ld_G(q, qd) = dot(J(q), qdd) + dot(qd.T, dot(H(q), qd)),
q -- joint angle values
"""
p_G = self.compute_com(q)
return self.compute_am_hessian(q, p_G, dof_indices)
def compute_com_hessian(self, q, dof_indices=None):
Hcom = zeros((self.nb_dofs, 3, self.nb_dofs))
with self.rave:
self.set_dof_values(q, dof_indices)
for link in self.rave.GetLinks():
index = link.GetIndex()
com = link.GetGlobalCOM()
m = link.GetMass()
H = self.rave.ComputeHessianTranslation(index, com)
Hcom += m * H
H = Hcom / self.mass
if dof_indices:
return ((H[dof_indices, :, :])[:, :, dof_indices])
elif self.active_dofs and len(q) == self.nb_active_dofs:
return ((H[self.active_dofs, :, :])[:, :, self.active_dofs])
return H
def compute_link_hessian(self, link, q, dof_indices=None):
with self.rave:
self.set_dof_values(q, dof_indices)
H_trans = self.rave.ComputeHessianTranslation(link.index, link.p)
H_rot = self.rave.ComputeHessianAxisAngle(link.index)
H = hstack([H_rot, H_trans])
if dof_indices:
return ((H[dof_indices, :, :])[:, :, dof_indices])
elif self.active_dofs and len(q) == self.nb_active_dofs:
return ((H[self.active_dofs, :, :])[:, :, self.active_dofs])
return H
def init_ik(self, dt, qd_lim, K_doflim, doflim_scale=0.95):
self.ik = DiffIKSolver(self, dt, qd_lim, K_doflim, doflim_scale)
