check time validity before pos validity

This commit is contained in:
Hongyi Zhou 2022-11-08 13:57:32 +01:00
parent d384e6e764
commit b9c2348855
3 changed files with 109 additions and 83 deletions

View File

@ -164,95 +164,101 @@ class BlackBoxWrapper(gym.ObservationWrapper):
def step(self, action: np.ndarray):
""" This function generates a trajectory based on a MP and then does the usual loop over reset and step"""
time_valid = self.env.check_time_validity(action)
if self.plan_counts == 0:
self.tau_first_prediction = action[0]
if time_valid:
## tricky part, only use weights basis
# basis_weights = action.reshape(7, -1)
# goal_weights = np.zeros((7, 1))
# action = np.concatenate((basis_weights, goal_weights), axis=1).flatten()
if self.plan_counts == 0:
self.tau_first_prediction = action[0]
# TODO remove this part, right now only needed for beer pong
# mp_params, env_spec_params, proceed = self.env.episode_callback(action, self.traj_gen)
position, velocity = self.get_trajectory(action)
traj_is_valid = self.env.episode_callback(action, position, velocity)
## tricky part, only use weights basis
# basis_weights = action.reshape(7, -1)
# goal_weights = np.zeros((7, 1))
# action = np.concatenate((basis_weights, goal_weights), axis=1).flatten()
trajectory_length = len(position)
rewards = np.zeros(shape=(trajectory_length,))
if self.verbose >= 2:
actions = np.zeros(shape=(trajectory_length,) + self.env.action_space.shape)
observations = np.zeros(shape=(trajectory_length,) + self.env.observation_space.shape,
dtype=self.env.observation_space.dtype)
# TODO remove this part, right now only needed for beer pong
# mp_params, env_spec_params, proceed = self.env.episode_callback(action, self.traj_gen)
position, velocity = self.get_trajectory(action)
traj_is_valid = self.env.episode_callback(action, position, velocity)
infos = dict()
done = False
trajectory_length = len(position)
rewards = np.zeros(shape=(trajectory_length,))
if self.verbose >= 2:
actions = np.zeros(shape=(trajectory_length,) + self.env.action_space.shape)
observations = np.zeros(shape=(trajectory_length,) + self.env.observation_space.shape,
dtype=self.env.observation_space.dtype)
if self.verbose >= 2:
desired_pos_traj = []
desired_vel_traj = []
pos_traj = []
vel_traj = []
if traj_is_valid:
self.plan_counts += 1
for t, (pos, vel) in enumerate(zip(position, velocity)):
step_action = self.tracking_controller.get_action(pos, vel, self.current_pos, self.current_vel)
c_action = np.clip(step_action, self.env.action_space.low, self.env.action_space.high)
obs, c_reward, done, info = self.env.step(c_action)
rewards[t] = c_reward
if self.verbose >= 2:
actions[t, :] = c_action
observations[t, :] = obs
for k, v in info.items():
elems = infos.get(k, [None] * trajectory_length)
elems[t] = v
infos[k] = elems
if self.verbose >= 2:
desired_pos_traj.append(pos)
desired_vel_traj.append(vel)
pos_traj.append(self.current_pos)
vel_traj.append(self.current_vel)
if self.render_kwargs:
self.env.render(**self.render_kwargs)
if done or self.replanning_schedule(self.current_pos, self.current_vel, obs, c_action,
t + 1 + self.current_traj_steps):
# if self.max_planning_times is not None and self.plan_counts >= self.max_planning_times:
# continue
self.condition_pos = pos if self.desired_conditioning else self.current_pos
self.condition_vel = vel if self.desired_conditioning else self.current_vel
break
infos.update({k: v[:t+1] for k, v in infos.items()})
self.current_traj_steps += t + 1
infos = dict()
done = False
if self.verbose >= 2:
infos['desired_pos'] = position[:t+1]
infos['desired_vel'] = velocity[:t+1]
infos['current_pos'] = self.current_pos
infos['current_vel'] = self.current_vel
infos['step_actions'] = actions[:t + 1]
infos['step_observations'] = observations[:t + 1]
infos['step_rewards'] = rewards[:t + 1]
infos['desired_pos_traj'] = np.array(desired_pos_traj)
infos['desired_vel_traj'] = np.array(desired_vel_traj)
infos['pos_traj'] = np.array(pos_traj)
infos['vel_traj'] = np.array(vel_traj)
desired_pos_traj = []
desired_vel_traj = []
pos_traj = []
vel_traj = []
infos['trajectory_length'] = t + 1
trajectory_return = self.reward_aggregation(rewards[:t + 1])
return self.observation(obs), trajectory_return, done, infos
if traj_is_valid:
self.plan_counts += 1
for t, (pos, vel) in enumerate(zip(position, velocity)):
step_action = self.tracking_controller.get_action(pos, vel, self.current_pos, self.current_vel)
c_action = np.clip(step_action, self.env.action_space.low, self.env.action_space.high)
obs, c_reward, done, info = self.env.step(c_action)
rewards[t] = c_reward
if self.verbose >= 2:
actions[t, :] = c_action
observations[t, :] = obs
for k, v in info.items():
elems = infos.get(k, [None] * trajectory_length)
elems[t] = v
infos[k] = elems
if self.verbose >= 2:
desired_pos_traj.append(pos)
desired_vel_traj.append(vel)
pos_traj.append(self.current_pos)
vel_traj.append(self.current_vel)
if self.render_kwargs:
self.env.render(**self.render_kwargs)
if done or self.replanning_schedule(self.current_pos, self.current_vel, obs, c_action,
t + 1 + self.current_traj_steps):
# if self.max_planning_times is not None and self.plan_counts >= self.max_planning_times:
# continue
self.condition_pos = pos if self.desired_conditioning else self.current_pos
self.condition_vel = vel if self.desired_conditioning else self.current_vel
break
infos.update({k: v[:t+1] for k, v in infos.items()})
self.current_traj_steps += t + 1
if self.verbose >= 2:
infos['desired_pos'] = position[:t+1]
infos['desired_vel'] = velocity[:t+1]
infos['current_pos'] = self.current_pos
infos['current_vel'] = self.current_vel
infos['step_actions'] = actions[:t + 1]
infos['step_observations'] = observations[:t + 1]
infos['step_rewards'] = rewards[:t + 1]
infos['desired_pos_traj'] = np.array(desired_pos_traj)
infos['desired_vel_traj'] = np.array(desired_vel_traj)
infos['pos_traj'] = np.array(pos_traj)
infos['vel_traj'] = np.array(vel_traj)
infos['trajectory_length'] = t + 1
trajectory_return = self.reward_aggregation(rewards[:t + 1])
return self.observation(obs), trajectory_return, done, infos
else:
obs, trajectory_return, done, infos = self.env.invalid_traj_callback(action, position, velocity)
return self.observation(obs), trajectory_return, done, infos
else:
obs, trajectory_return, done, infos = self.env.invalid_traj_callback(action, position, velocity)
return self.observation(obs), trajectory_return, done, infos
obs, trajectory_return, done, infos = self.env.time_invalid_traj_callback(action)
return self.observation(obs), trajectory_return, done, infos
def render(self, **kwargs):
"""Only set render options here, such that they can be used during the rollout.
This only needs to be called once"""

View File

@ -560,7 +560,7 @@ for _v in _versions:
_name = _v.split("-")
_env_id = f'{_name[0]}ProDMP-{_name[1]}'
kwargs_dict_tt_prodmp = deepcopy(DEFAULT_BB_DICT_ProDMP)
kwargs_dict_tt_prodmp['wrappers'].append(mujoco.box_pushing.MPWrapper)
kwargs_dict_tt_prodmp['wrappers'].append(mujoco.table_tennis.MPWrapper)
kwargs_dict_tt_prodmp['name'] = _v
kwargs_dict_tt_prodmp['controller_kwargs']['p_gains'] = 0.5 * np.array([1.0, 4.0, 2.0, 4.0, 1.0, 4.0, 1.0])
kwargs_dict_tt_prodmp['controller_kwargs']['d_gains'] = 0.5 * np.array([0.1, 0.4, 0.2, 0.4, 0.1, 0.4, 0.1])

View File

@ -28,12 +28,31 @@ class MPWrapper(RawInterfaceWrapper):
def current_vel(self) -> Union[float, int, np.ndarray, Tuple]:
return self.data.qvel[:7].copy()
def check_time_validity(self, action):
return action[0] <= tau_bound[1] and action[0] >= tau_bound[0] \
and action[1] <= delay_bound[1] and action[1] >= delay_bound[0]
def time_invalid_traj_callback(self, action) \
-> Tuple[np.ndarray, float, bool, dict]:
tau_invalid_penalty = 3 * (np.max([0, action[0] - tau_bound[1]]) + np.max([0, tau_bound[0] - action[0]]))
delay_invalid_penalty = 3 * (np.max([0, action[1] - delay_bound[1]]) + np.max([0, delay_bound[0] - action[1]]))
invalid_penalty = tau_invalid_penalty + delay_invalid_penalty
obs = np.concatenate([self.get_obs(), np.array([0])])
return obs, -invalid_penalty, True, {
"hit_ball": [False],
"ball_returned_success": [False],
"land_dist_error": [10.],
"is_success": [False],
'trajectory_length': 1,
"num_steps": [1]
}
def episode_callback(self, action, pos_traj, vel_traj):
time_invalid = action[0] > tau_bound[1] or action[0] < tau_bound[0] \
or action[1] > delay_bound[1] or action[1] < delay_bound[0]
if time_invalid or np.any(pos_traj > jnt_pos_high) or np.any(pos_traj < jnt_pos_low):
return False
return True
return True
return False
def invalid_traj_callback(self, action, pos_traj: np.ndarray, vel_traj: np.ndarray) \
-> Tuple[np.ndarray, float, bool, dict]:
@ -43,7 +62,8 @@ class MPWrapper(RawInterfaceWrapper):
violate_low_bound_error = np.mean(np.maximum(jnt_pos_low - pos_traj, 0))
invalid_penalty = tau_invalid_penalty + delay_invalid_penalty + \
violate_high_bound_error + violate_low_bound_error
return self.get_obs(), -invalid_penalty, True, {
obs = np.concatenate([self.get_obs(), np.array([0])])
return obs, -invalid_penalty, True, {
"hit_ball": [False],
"ball_returned_success": [False],
"land_dist_error": [10.],