def step(self, ctrl):
ctrl = self.scale_action(ctrl)
self.sim.data.ctrl[:] = ctrl
self.sim.forward()
self.sim.step()
self.step_ctr += 1
obs = self.get_obs()
# Angle deviation
x, y, z, qw, qx, qy, qz = obs[:7]
xd, yd, zd, thd, phid, psid = self.sim.get_state().qvel.tolist()[:6]
# Reward conditions
target_vel = 0.25
t_x = target_vel * np.cos(self.target_yaw)
t_y = target_vel * np.sin(self.target_yaw)
velocity_rew_x = 1. / (abs(xd - t_x) + 1.) - 1. / (t_x + 1.)
velocity_rew_y = 1. / (abs(yd - t_y) + 1.) - 1. / (t_y + 1.)
roll, pitch, yaw = my_utils.quat_to_rpy([qw, qx, qy, qz])
yaw_pen = np.min((abs((yaw % 6.183) - (self.target_yaw % 6.183)),
abs(yaw - self.target_yaw)))
r_pos = velocity_rew_x * 6 + velocity_rew_y * 6
r_neg = np.square(roll) * 0.5 + \
np.square(pitch) * 0.5 + \
np.square(zd) * 0.5 + \
np.square(yaw_pen) * 0.2
r_neg = np.clip(r_neg, 0, 1) * 1.
r_pos = np.clip(r_pos, -2, 2)
r = r_pos - r_neg
self.episode_reward += r
# Reevaluate termination condition
done = self.step_ctr > self.max_steps
contacts = (np.abs(
np.array(self.sim.data.cfrc_ext[[4, 7, 10, 13, 16, 19
]])).sum(axis=1) > 0.05).astype(
np.float32)
obs = np.concatenate([
self.scale_joints(self.sim.get_state().qpos.tolist()[7:]),
self.sim.get_state().qvel.tolist()[6:],
self.sim.get_state().qvel.tolist()[:6], [roll, pitch, yaw, y],
[self.target_yaw], contacts
])
self.model.body_quat[20] = my_utils.rpy_to_quat(0, 0, self.target_yaw)
self.model.body_pos[20] = [x, y, 1]
return obs, r, done, None
def step(self, ctrl):
ctrl = self.scale_action(ctrl)
self.sim.data.ctrl[:] = ctrl
self.sim.forward()
self.sim.step()
self.step_ctr += 1
obs = self.get_obs()
obs_dict = self.get_obs_dict()
# Angle deviation
x, y, z, qw, qx, qy, qz = obs[:7]
xd, yd, zd, thd, phid, psid = self.sim.get_state().qvel.tolist()[:6]
contacts = (np.abs(
np.array(self.sim.data.cfrc_ext[[4, 7, 10, 13, 16, 19
]])).sum(axis=1) > 0.05).astype(
np.float32)
# Reward conditions
target_vel = self.target_criteria_dict["vel"]
velocity_rew = (1. / (abs(xd - target_vel) + 1.) - 1. /
(target_vel + 1.)) * (1. / target_vel)
roll, pitch, yaw = my_utils.quat_to_rpy([qw, qx, qy, qz])
r_pos = velocity_rew
r_neg = np.square(self.target_criteria_dict["height"] - z) * 100 + \
np.square(roll) * 1. + \
np.square(pitch) * 1. + \
np.square(zd) * 1. + \
np.square(yaw) * 5. + \
np.square(y) * 1.
#print(z, self.target_criteria_dict["height"], np.square(self.target_criteria_dict["height"] - z) * 100, velocity_rew)
r_neg = np.clip(r_neg, 0, 2.)
r_pos = np.clip(r_pos, -2, 2)
r = r_pos - r_neg
# Reevaluate termination condition
done = self.step_ctr > self.max_steps # or abs(roll) > 2 or abs(pitch) > 2
obs = np.concatenate([
self.scale_joints(self.sim.get_state().qpos.tolist()[7:]),
self.sim.get_state().qvel.tolist()[6:],
self.sim.get_state().qvel.tolist()[:6], self.target_criteria_norm,
[roll, pitch, yaw, y], contacts
])
return obs, r, done, obs_dict
def step(self, ctrl):
ctrl = self.scale_action(ctrl)
self.sim.data.ctrl[:] = ctrl
self.sim.forward()
self.sim.step()
self.step_ctr += 1
obs = self.get_obs()
# Angle deviation
x, y, z, qw, qx, qy, qz = obs[:7]
xd, yd, zd, thd, phid, psid = self.sim.get_state().qvel.tolist()[:6]
# Reward conditions
target_vel = 0.30
velocity_rew = 1. / (abs(xd - target_vel) + 1.) - 1. / (target_vel +
1.)
roll, pitch, yaw = my_utils.quat_to_rpy([qw, qx, qy, qz])
r_pos = velocity_rew * 5
r_neg = np.square(self.sim.data.actuator_force).mean() * 0.00001 + \
np.square(ctrl).mean() * 0.01 + \
np.square(roll) * 0.1 + \
np.square(pitch) * 0.1 + \
np.square(yaw) * 0.3 + \
np.square(y) * 0.1
r_neg = np.clip(r_neg, 0, 1) * 1
r_pos = np.clip(r_pos, -2, 2)
r = r_pos - r_neg
self.episode_reward += r
# Reevaluate termination condition
done = self.step_ctr > self.max_steps
contacts = (np.abs(
np.array(self.sim.data.cfrc_ext[[4, 7, 10, 13, 16, 19
]])).sum(axis=1) > 0.05).astype(
np.float32)
obs = np.concatenate([
self.sim.get_state().qpos.tolist()[7:],
self.sim.get_state().qvel.tolist()[6:],
self.sim.get_state().qvel.tolist()[:6], [roll, pitch, yaw, y],
contacts
])
return obs, r, done, None
def step(self, ctrl):
goal_dist_pre = np.abs(self.goal_pos -
self.sim.get_state().qpos.tolist()[:2]).sum()
mem = ctrl[-self.mem_dim:]
act = ctrl[:-self.mem_dim]
sact = self.scale_action(act)
self.sim.data.ctrl[:] = sact
self.sim.step()
self.step_ctr += 1
goal_dist_post = np.abs(self.goal_pos -
self.sim.get_state().qpos.tolist()[:2]).sum()
obs_c = self.get_obs()
obs_dict = self.get_obs_dict()
x, y, z, qw, qx, qy, qz = obs_c[:7]
#angle = 2 * np.arccos(qw)
_, _, yaw = my_utils.quat_to_rpy((qw, qx, qy, qz))
target_angle = np.arctan2(self.goal_pos[1] - y, self.goal_pos[0] - x)
target_err = abs(yaw - target_angle)
# Reevaluate termination condition.
done = self.step_ctr > self.max_steps # or obs_dict['torso_contact'] > 0.1
xd, yd, _, _, _, _ = obs_dict["root_vel"]
ctrl_effort = np.square(ctrl[0:8]).mean() * 0.00
target_progress = (goal_dist_pre - goal_dist_post) * 50
#print(target_progress)
obs = np.concatenate(
(self.goal_pos - obs_c[:2], obs_c[2:], obs_dict["contacts"],
[obs_dict['torso_contact']], mem))
r = np.clip(target_progress, -1,
1) - ctrl_effort - min(target_err, 3) * 0.1
return obs, r, done, obs_dict
def step(self, ctrl):
ctrl = self.scale_action(ctrl)
self.sim.data.ctrl[:] = ctrl
self.sim.forward()
self.sim.step()
self.step_ctr += 1
obs = self.get_obs()
obs_dict = self.get_obs_dict()
# Angle deviation
x, y, z, qw, qx, qy, qz = obs[:7]
xd, yd, zd, thd, phid, psid = self.sim.get_state().qvel.tolist()[:6]
# Reward conditions
ctrl_effort = np.square(ctrl).sum()
target_vel = 0.25
velocity_rew = 1. / (abs(xd - target_vel) + 1.) - 1. / (target_vel +
1.)
roll, pitch, yaw = my_utils.quat_to_rpy([qw, qx, qy, qz])
r = velocity_rew * 5 - \
np.square(self.sim.data.actuator_force).mean() * 0.0001
r = np.clip(r, -2, 2)
self.episode_reward += r
# Reevaluate termination condition
done = self.step_ctr > self.max_steps #or abs(roll) > 2 or abs(pitch) > 2
obs = np.concatenate([
np.array(self.sim.get_state().qpos.tolist()[7:]),
[roll, pitch, yaw, xd, yd, thd, phid], obs_dict["contacts"]
])
return obs, r, done, obs_dict
def step(self, ctrl):
ctrl = np.clip(ctrl, -1, 1)
ctrl_pen = np.square(ctrl).mean()
#torques = self.sim.data.actuator_force
#ctrl_pen = np.square(torques).mean()
ctrl = self.scale_action(ctrl)
self.sim.data.ctrl[:] = ctrl
self.sim.forward()
self.sim.step()
self.step_ctr += 1
#
# if np.random.rand() < 0.002:
# self.model.geom_friction[0] = np.array([0.0001, 0.5, 0.5])
# print("Changed friction")
obs = self.get_obs()
# Angle deviation
x, y, z, qw, qx, qy, qz = obs[:7]
xd, yd, zd, thd, phid, psid = self.sim.get_state().qvel.tolist()[:6]
#xa, ya, za, tha, phia, psia = self.sim.data.qacc.tolist()[:6]
self.vel_sum += xd
# Reward conditions
target_vel = 0.4
velocity_rew = 1. / (abs(xd - target_vel) + 1.) - 1. / (target_vel +
1.)
velocity_rew = velocity_rew * (1 / (1 + 30 * np.square(yd)))
roll, pitch, _ = my_utils.quat_to_rpy([qw, qx, qy, qz])
#yaw_deviation = np.min((abs((yaw % 6.183) - (0 % 6.183)), abs(yaw - 0)))
q_yaw = 2 * acos(qw)
r_neg = np.square(y) * 0.2 + \
np.square(q_yaw) * 0.5 + \
np.square(pitch) * 0.5 + \
np.square(roll) * 0.5 + \
ctrl_pen * 0.01 + \
np.square(zd) * 0.7
r_pos = velocity_rew * 6 # + (abs(self.prev_yaw_deviation) - abs(yaw_deviation)) * 3. + (abs(self.prev_y_deviation) - abs(y)) * 3.
r = r_pos - r_neg
# Reevaluate termination condition
done = self.step_ctr > self.max_steps # or abs(y) > 0.3 or abs(yaw) > 0.6 or abs(roll) > 0.8 or abs(pitch) > 0.8
contacts = (np.abs(
np.array(self.sim.data.cfrc_ext[[4, 7, 10, 13, 16, 19
]])).sum(axis=1) > 0.05).astype(
np.float32)
if self.use_HF:
obs = np.concatenate([
self.scale_joints(self.sim.get_state().qpos.tolist()[7:]),
self.sim.get_state().qvel.tolist()[6:],
self.sim.get_state().qvel.tolist()[:6], [qw, qx, qy, qz, y],
contacts,
self.get_local_hf(x, y).flatten()
])
else:
obs = np.concatenate([
self.scale_joints(self.sim.get_state().qpos.tolist()[7:]),
self.sim.get_state().qvel.tolist()[6:],
self.sim.get_state().qvel.tolist()[:6], [qw, qx, qy, qz, y],
contacts
])
return obs, r, done, None
def step(self, ctrl):
ctrl = self.scale_action(ctrl)
self.sim.data.ctrl[:] = ctrl
self.sim.forward()
self.sim.step()
self.step_ctr += 1
obs = self.get_obs()
# Angle deviation
x, y, z, qw, qx, qy, qz = obs[:7]
xd, yd, zd, thd, phid, psid = self.sim.get_state().qvel.tolist()[:6]
# Reward conditions
target_vel = 0.25
dprev = np.sqrt((self.prev_xy[0] - self.goal_xy[0])**2 +
(self.prev_xy[1] - self.goal_xy[1])**2)
dcur = np.sqrt((x - self.goal_xy[0])**2 + (y - self.goal_xy[1])**2)
to_goal_vel = dprev - dcur
to_goal_vel *= 50.
velocity_rew = 1. / (abs(to_goal_vel - target_vel) +
1.) - 1. / (target_vel + 1.)
roll, pitch, yaw = my_utils.quat_to_rpy([qw, qx, qy, qz])
tar_angle = np.arctan2(self.goal_xy[1] - y, self.goal_xy[0] - x)
yaw_deviation = np.min(
(abs((yaw % 6.183) - (tar_angle % 6.183)), abs(yaw - tar_angle)))
r_pos = velocity_rew * 3 + (self.prev_deviation - yaw_deviation) * 9
r_neg = np.square(roll) * 1.2 + \
np.square(pitch) * 1.2 + \
np.square(zd) * 1.5
self.prev_deviation = yaw_deviation
r_neg = np.clip(r_neg, 0, 2) * 1.
r_pos = np.clip(r_pos, -2, 2)
r = r_pos - r_neg
self.episode_reward += r
# Reevaluate termination condition
done = self.step_ctr > self.max_steps or dcur < 0.15
self.prev_xy = [x, y]
contacts = (np.abs(
np.array(self.sim.data.cfrc_ext[[4, 7, 10, 13, 16, 19
]])).sum(axis=1) > 0.05).astype(
np.float32)
obs = np.concatenate([
self.scale_joints(self.sim.get_state().qpos.tolist()[7:]),
self.sim.get_state().qvel.tolist()[6:],
self.sim.get_state().qvel.tolist()[:6], [roll, pitch, yaw, y],
[x - self.goal_xy[0], y - self.goal_xy[1]], contacts
])
return obs, r, done, None
def step(self, ctrl):
ctrl = ctrl[:-self.n_envs]
classif = ctrl[-self.n_envs:]
ctrl = np.clip(ctrl, -1, 1)
ctrl = self.scale_action(ctrl)
self.sim.data.ctrl[:] = ctrl
self.sim.forward()
self.sim.step()
self.step_ctr += 1
obs = self.get_obs()
# Angle deviation
x, y, z, qw, qx, qy, qz = obs[:7]
xd, yd, zd, thd, phid, psid = self.sim.get_state().qvel.tolist()[:6]
xa, ya, za, tha, phia, psia = self.sim.data.qacc.tolist()[:6]
if x > self.scaled_indeces_list[self.current_env_idx]:
self.current_env_idx += 1
self.current_env = self.envs[self.current_env_idx]
with T.no_grad():
env_label = T.tensor(self.env_list.index(
self.current_env)).unsqueeze(0)
env_pred = T.tensor(classif).unsqueeze(0)
ce_loss = self.CE_loss(env_pred, env_label)
# Reward conditions
target_vel = 0.3
velocity_rew = 1. / (abs(xd - target_vel) + 1.) - 1. / (target_vel +
1.)
roll, pitch, yaw = my_utils.quat_to_rpy([qw, qx, qy, qz])
r_pos = velocity_rew * 6
r_neg = np.square(roll) * 2. + \
np.square(pitch) * 2. + \
np.square(zd) * 2. + \
np.square(yd) * 2. + \
np.square(y) * 2. + \
np.square(yaw) * 4.0 + \
np.square(self.sim.data.actuator_force).mean() * 0.000 + \
np.clip(np.square(np.array(self.sim.data.cfrc_ext[1])).sum(axis=0), 0, 1) * 0.4 + \
ce_loss.numpy() * .1
r_neg = np.clip(r_neg, 0, 1) * 1
r_pos = np.clip(r_pos, -2, 2)
r = r_pos - r_neg
self.episode_reward += r
# Reevaluate termination condition
done = self.step_ctr > self.max_steps
contacts = (np.abs(
np.array(self.sim.data.cfrc_ext[[4, 7, 10, 13, 16, 19
]])).sum(axis=1) > 0.05).astype(
np.float32)
if self.use_HF:
obs = np.concatenate([
self.scale_joints(self.sim.get_state().qpos.tolist()[7:]),
self.sim.get_state().qvel.tolist()[6:],
self.sim.get_state().qvel.tolist()[:6], [roll, pitch, yaw, y],
contacts,
self.get_local_hf(x, y).flatten()
])
else:
obs = np.concatenate([
self.scale_joints(self.sim.get_state().qpos.tolist()[7:]),
self.sim.get_state().qvel.tolist()[6:],
self.sim.get_state().qvel.tolist()[:6], [roll, pitch, yaw, y],
contacts
])
return obs, r, done, None
def step(self, ctrl):
ctrl = np.clip(ctrl, -1, 1)
ctrl = self.scale_action(ctrl)
self.sim.data.ctrl[:] = ctrl
self.sim.forward()
self.sim.step()
self.step_ctr += 1
obs = self.get_obs()
# Angle deviation
x, y, z, qw, qx, qy, qz = obs[:7]
xd, yd, zd, thd, phid, psid = self.sim.get_state().qvel.tolist()[:6]
if x > self.scaled_indeces_list[self.current_env_idx]:
self.current_env_idx += 1
self.current_env = self.envs[self.current_env_idx]
expert = self.expert_dict[self.current_env]
expert_act = expert(my_utils.to_tensor(self.obs, True)).detach()
expert_act = expert_act[0].numpy()
# Reward conditions
target_vel = 0.3
velocity_rew = 1. / (abs(xd - target_vel) + 1.) - 1. / (target_vel +
1.)
roll, pitch, yaw = my_utils.quat_to_rpy([qw, qx, qy, qz])
r_pos = velocity_rew * 6
r_neg = np.square(roll) * 1. + \
np.square(pitch) * 1. + \
np.square(zd) * 1. + \
np.square(y) * 1. + \
np.square(yaw) * 1.0 + \
np.square(ctrl - expert_act).mean() * 0.01
print(np.square(ctrl - expert_act).mean() * 0.01)
r_neg = np.clip(r_neg, 0, 1) * 1
r_pos = np.clip(r_pos, -2, 2)
r = r_pos - r_neg
self.episode_reward += r
# Reevaluate termination condition
done = self.step_ctr > self.max_steps
contacts = (np.abs(
np.array(self.sim.data.cfrc_ext[[4, 7, 10, 13, 16, 19
]])).sum(axis=1) > 0.05).astype(
np.float32)
if self.use_HF:
obs = np.concatenate([
self.scale_joints(self.sim.get_state().qpos.tolist()[7:]),
self.sim.get_state().qvel.tolist()[6:],
self.sim.get_state().qvel.tolist()[:6], [roll, pitch, yaw, y],
contacts,
self.get_local_hf(x, y).flatten()
])
else:
obs = np.concatenate([
self.scale_joints(self.sim.get_state().qpos.tolist()[7:]),
self.sim.get_state().qvel.tolist()[6:],
self.sim.get_state().qvel.tolist()[:6], [roll, pitch, yaw, y],
contacts
])
self.obs = obs
return obs, r, done, self.env_list.index(self.current_env)
def step(self, ctrl):
ctrl = ctrl[:-(self.n_envs + 1)]
classif = ctrl[-(self.n_envs + 1):]
if np.argmax(classif) < self.n_envs + 1:
policy = self.expert_list[np.argmax(classif[:self.n_envs])]
act = policy(my_utils.to_tensor(self.obs, True)).detach()
ctrl = act[0].numpy()
ctrl = np.clip(ctrl, -1, 1)
ctrl = self.scale_action(ctrl)
self.sim.data.ctrl[:] = ctrl
self.sim.forward()
self.sim.step()
self.step_ctr += 1
obs = self.get_obs()
# Angle deviation
x, y, z, qw, qx, qy, qz = obs[:7]
xd, yd, zd, thd, phid, psid = self.sim.get_state().qvel.tolist()[:6]
xa, ya, za, tha, phia, psia = self.sim.data.qacc.tolist()[:6]
# Reward conditions
target_vel = 0.3
velocity_rew = 1. / (abs(xd - target_vel) + 1.) - 1. / (target_vel +
1.)
roll, pitch, yaw = my_utils.quat_to_rpy([qw, qx, qy, qz])
r_pos = velocity_rew * 6
r_neg = np.square(roll) * 2. + \
np.square(pitch) * 2. + \
np.square(zd) * 2. + \
np.square(yd) * 2. + \
np.square(y) * 2. + \
np.square(yaw) * 4.0 + \
np.square(self.sim.data.actuator_force).mean() * 0.000 + \
np.clip(np.square(np.array(self.sim.data.cfrc_ext[1])).sum(axis=0), 0, 1) * 0.4
r_neg = np.clip(r_neg, 0, 1) * 1
r_pos = np.clip(r_pos, -2, 2)
r = r_pos - r_neg
self.episode_reward += r
# Reevaluate termination condition
done = self.step_ctr > self.max_steps
contacts = (np.abs(
np.array(self.sim.data.cfrc_ext[[4, 7, 10, 13, 16, 19
]])).sum(axis=1) > 0.05).astype(
np.float32)
if self.use_HF:
self.obs = np.concatenate([
self.scale_joints(self.sim.get_state().qpos.tolist()[7:]),
self.sim.get_state().qvel.tolist()[6:],
self.sim.get_state().qvel.tolist()[:6], [roll, pitch, yaw, y],
contacts,
self.get_local_hf(x, y).flatten()
])
else:
self.obs = np.concatenate([
self.scale_joints(self.sim.get_state().qpos.tolist()[7:]),
self.sim.get_state().qvel.tolist()[6:],
self.sim.get_state().qvel.tolist()[:6], [roll, pitch, yaw, y],
contacts
])
return self.obs, r, done, None
def step(self, ctrl):
# Mute appropriate leg joints
for i in range(6):
if self.dead_leg_vector[i] == 1:
ctrl[i * 3:i * 3 + 3] = np.zeros(3) #np.random.randn(3) * 0.1
if self.mem_dim == 0:
mem = np.zeros(0)
act = ctrl
ctrl = self.scale_action(act)
else:
mem = ctrl[-self.mem_dim:]
act = ctrl[:-self.mem_dim]
ctrl = self.scale_action(act)
self.prev_act = np.array(act)
self.sim.data.ctrl[:] = ctrl
self.sim.forward()
self.sim.step()
self.step_ctr += 1
obs = self.get_obs()
obs_dict = self.get_obs_dict()
# Angle deviation
x, y, z, qw, qx, qy, qz = obs[:7]
xd, yd, zd, _, _, _ = self.sim.get_state().qvel.tolist()[:6]
angle = 2 * acos(qw)
roll, pitch, yaw = my_utils.quat_to_rpy((qw, qx, qy, qz))
# Reward conditions
target_vel = 0.25
velocity_rew = 1. / (abs(xd - target_vel) + 1.) - 1. / (target_vel +
1.)
r = velocity_rew * 10 - \
np.square(self.sim.data.actuator_force).mean() * 0.0001 - \
np.abs(roll) * 0.1 - \
np.square(pitch) * 0.1 - \
np.square(yaw) * .1 - \
np.square(y) * 0.1 - \
np.square(zd) * 0.01
r = np.clip(r, -2, 2)
self.cumulative_environment_reward += r
# Reevaluate termination condition
done = self.step_ctr > self.max_steps # or abs(y) > 0.3 or x < -0.2 or abs(yaw) > 0.8
obs = np.concatenate([
np.array(self.sim.get_state().qpos.tolist()[3:]), [xd, yd],
obs_dict["contacts"], mem
])
# if np.random.rand() < self.dead_leg_prob:
# idx = np.random.randint(0,6)
# self.dead_leg_vector[idx] = 1
# self.dead_leg_sums[idx] += 1
# self.model.geom_rgba[self.model._geom_name2id[self.leg_list[idx]]] = [1, 0, 0, 1]
# self.dead_leg_prob = 0.
return obs, r, done, obs_dict
def step(self, ctrl):
ctrl_penalty = np.mean(ctrl)
ctrl = self.scale_action(ctrl)
self.sim.data.ctrl[:] = ctrl
self.sim.forward()
self.sim.step()
self.step_ctr += 1
obs = self.get_obs()
self.used_energy += self.sim.data.actuator_force
# Angle deviation
x, y, z, qw, qx, qy, qz = obs[:7]
xd, yd, zd, thd, phid, psid = self.sim.get_state().qvel.tolist()[:6]
# Reward conditions
target_vel = 0.30
velocity_rew_x = 1. / (abs(xd - target_vel) + 1.) - 1. / (target_vel +
1.)
roll, pitch, yaw = my_utils.quat_to_rpy([qw, qx, qy, qz])
contacts = (np.abs(
np.array(self.sim.data.cfrc_ext[[4, 7, 10, 13, 16, 19
]])).sum(axis=1) > 0.05).astype(
np.float32)
# mean_used_energy = self.used_energy / self.step_ctr
# mean_used_energy_coxa = mean_used_energy[0::3]
# mean_used_energy_femur = mean_used_energy[1::3]
# mean_used_energy_tibia = mean_used_energy[2::3]
# coxa_penalty = np.mean(np.square(mean_used_energy_coxa - np.mean(mean_used_energy_coxa)))
# femur_penalty = np.mean(np.square(mean_used_energy_femur - np.mean(mean_used_energy_femur)))
# tibia_penalty = np.mean(np.square(mean_used_energy_tibia - np.mean(mean_used_energy_tibia)))
r_pos = velocity_rew_x * 5. # + np.mean(contacts) * 0.2
r_neg = np.square(roll) * 1. + \
np.square(pitch) * 1. + \
np.square(zd) * 1. + \
np.square(yd) * 4. + \
np.square(y) * 7. + \
np.square(yaw) * 4. + \
np.square(self.sim.data.actuator_force).mean() * 0.000 + \
ctrl_penalty * 0.0 # + \
#(coxa_penalty * .1 + femur_penalty * .1 + tibia_penalty * .1) * self.step_ctr > 30
r_neg = np.clip(r_neg, 0, 1) * 1.
r_pos = np.clip(r_pos, -2, 2)
r = r_pos - r_neg
self.episode_reward += r
# Reevaluate termination condition
done = self.step_ctr > self.max_steps
obs = np.concatenate([
self.scale_joints(self.sim.get_state().qpos.tolist()[7:]),
self.sim.get_state().qvel.tolist()[6:],
self.sim.get_state().qvel.tolist()[:6], [roll, pitch, yaw, y],
contacts
])
return obs, r, done, None
def step(self, ctrl, render=False):
ctrl = np.clip(ctrl, -1, 1)
ctrl = self.scale_action(ctrl)
self.sim.data.ctrl[:] = ctrl
self.model.opt.timestep = 0.003
for i in range(27):
self.sim.forward()
self.sim.step()
if render:
self.render()
# Simulate read delay
self.model.opt.timestep = 0.0008
joints = []
for i in range(18):
joints.append(self.sim.get_state().qpos.tolist()[7 + i])
self.sim.forward()
self.sim.step()
self.step_ctr += 1
self.current_difficulty = np.minimum(
self.current_difficulty + self.difficulty_per_step_increment, 1)
torques = self.sim.data.actuator_force
ctrl_pen = np.square(torques).mean()
obs = self.get_obs()
# Angle deviation
x, y, z, qw, qx, qy, qz = obs[:7]
xd, yd, zd, thd, phid, psid = self.sim.get_state().qvel.tolist()[:6]
self.xd_queue.append(xd)
if len(self.xd_queue) > 15:
self.xd_queue.pop(0)
xd_av = sum(self.xd_queue) / len(self.xd_queue)
velocity_rew = 1. / (abs(xd_av - self.target_vel) +
1.) - 1. / (self.target_vel + 1.)
velocity_rew *= (0.3 / self.target_vel)
roll, pitch, _ = my_utils.quat_to_rpy([qw, qx, qy, qz])
q_yaw = 2 * acos(qw)
yaw_deviation = np.min(
(abs((q_yaw % 6.183) - (0 % 6.183)), abs(q_yaw - 0)))
r_neg = np.square(q_yaw) * 0.5 + \
np.square(pitch) * 1.5 * self.current_difficulty + \
np.square(roll) * 1.5 * self.current_difficulty + \
ctrl_pen * 0.00000 + \
np.square(zd) * 1.5 * self.current_difficulty
r_correction = np.clip(
abs(self.prev_deviation) - abs(yaw_deviation), -1, 1)
r_pos = velocity_rew * 7 + r_correction * 15
r = np.clip(r_pos - r_neg, -3, 3)
self.prev_deviation = yaw_deviation
# Reevaluate termination condition
done = self.step_ctr > self.max_steps
contacts = (np.abs(
np.array(self.sim.data.sensordata[0:6], dtype=np.float32)) >
0.05).astype(np.float32) * 2 - 1.
# See if changing contacts affects policy
#np.random.randint(0,2, size=6) * 2 - 1
#contacts = -np.ones(6)
#clipped_torques = np.clip(torques * 0.05, -1, 1)
c_obs = self.scale_joints(joints)
quat = obs[3:7]
if self.use_contacts:
c_obs = np.concatenate([c_obs, contacts])
c_obs = np.concatenate([c_obs, quat])
return c_obs, r, done, {}
def step(self, ctrl, render=False):
ctrl = np.clip(ctrl, -1, 1)
ctrl = self.scale_action(ctrl)
self.sim.data.ctrl[:] = ctrl
self.sim.forward()
self.sim.step()
if render:
self.render()
time.sleep(0.001)
self.step_ctr += 1
torques = self.sim.data.actuator_force
ctrl_pen = np.square(torques).mean()
obs = self.get_obs()
# Angle deviation
x, y, z, qw, qx, qy, qz = obs[:7]
xd, yd, zd, thd, phid, psid = self.sim.get_state().qvel.tolist()[:6]
self.xd_queue.append(xd)
if len(self.xd_queue) > 15:
self.xd_queue.pop(0)
xd_av = sum(self.xd_queue) / len(self.xd_queue)
velocity_rew = 1. / (abs(xd_av - self.target_vel) +
1.) - 1. / (self.target_vel + 1.)
velocity_rew *= (0.3 / self.target_vel)
roll, pitch, _ = my_utils.quat_to_rpy([qw, qx, qy, qz])
q_yaw = 2 * acos(qw)
yaw_deviation = np.min(
(abs((q_yaw % 6.183) - (0 % 6.183)), abs(q_yaw - 0)))
r_neg = np.square(q_yaw) * 0.2 + \
np.square(pitch) * 0.5 + \
np.square(roll) * 0.5 + \
ctrl_pen * 0.00001 + \
np.square(zd) * 0.7
r_correction = np.clip(
abs(self.prev_deviation) - abs(yaw_deviation), -1, 1)
r_pos = velocity_rew * 4 + r_correction * 10
r = np.clip(r_pos - r_neg, -3, 3)
self.prev_deviation = yaw_deviation
# Reevaluate termination condition
done = self.step_ctr > self.max_steps
contacts = (np.abs(
np.array(self.sim.data.sensordata[0:6], dtype=np.float32)) >
0.05).astype(np.float32) * 2 - 1.
#clipped_torques = np.clip(torques * 0.05, -1, 1)
c_obs = self.scale_joints(self.sim.get_state().qpos.tolist()[7:])
quat = obs[3:7]
if self.use_contacts:
c_obs = np.concatenate([c_obs, contacts])
c_obs = np.concatenate([c_obs, quat])
return np.array(quat), r, done, {}