Merge pull request #73 from HongyiZhouCN/69-mp-params-validity

1. The tau and delay bounds can be correctly set in init.py 
2. Add a random initialization flag to all box-pushing environments

.gitignore

@@ -106,6 +106,7 @@ venv.bak/
 # pycharm
 .DS_Store
 .idea
+.vscode
 
 #configs
 /configs/db.cfg

fancy_gym/black_box/black_box_wrapper.py

@@ -55,6 +55,14 @@ class BlackBoxWrapper(gym.ObservationWrapper):
         # self.traj_gen.set_mp_times(self.time_steps)
         self.traj_gen.set_duration(self.duration, self.dt)
 
+        # check
+        self.tau_bound = [-np.inf, np.inf]
+        self.delay_bound = [-np.inf, np.inf]
+        if hasattr(self.traj_gen.phase_gn, "tau_bound"):
+            self.tau_bound = self.traj_gen.phase_gn.tau_bound
+        if hasattr(self.traj_gen.phase_gn, "delay_bound"):
+            self.delay_bound = self.traj_gen.phase_gn.delay_bound
+
         # reward computation
         self.reward_aggregation = reward_aggregation
 
@@ -139,7 +147,8 @@ class BlackBoxWrapper(gym.ObservationWrapper):
 
         position, velocity = self.get_trajectory(action)
         position, velocity = self.env.set_episode_arguments(action, position, velocity)
-        traj_is_valid, position, velocity = self.env.preprocessing_and_validity_callback(action, position, velocity)
+        traj_is_valid, position, velocity = self.env.preprocessing_and_validity_callback(action, position, velocity,
+                                                                                         self.tau_bound, self.delay_bound)
 
         trajectory_length = len(position)
         rewards = np.zeros(shape=(trajectory_length,))
@@ -153,7 +162,8 @@ class BlackBoxWrapper(gym.ObservationWrapper):
 
         if not traj_is_valid:
             obs, trajectory_return, done, infos = self.env.invalid_traj_callback(action, position, velocity,
-                                                                                 self.return_context_observation)
+                                                                                 self.return_context_observation,
+                                                                                 self.tau_bound, self.delay_bound)
             return self.observation(obs), trajectory_return, done, infos
 
         self.plan_steps += 1

fancy_gym/black_box/raw_interface_wrapper.py

@@ -52,7 +52,8 @@ class RawInterfaceWrapper(gym.Wrapper):
         """
         return self.env.dt
 
-    def preprocessing_and_validity_callback(self, action: np.ndarray, pos_traj: np.ndarray, vel_traj: np.ndarray) \
+    def preprocessing_and_validity_callback(self, action: np.ndarray, pos_traj: np.ndarray, vel_traj: np.ndarray,
+                                            tau_bound: list = None, delay_bound: list = None ) \
             -> Tuple[bool, np.ndarray, np.ndarray]:
         """
         Used to preprocess the action and check if the desired trajectory is valid.
@@ -61,6 +62,8 @@ class RawInterfaceWrapper(gym.Wrapper):
             specified, else only traj_gen parameters
             pos_traj: a vector instance of the raw position trajectory
             vel_traj: a vector instance of the raw velocity trajectory
+            tau_bound: a list of two elements, the lower and upper bound of the trajectory length scaling factor
+            delay_bound: a list of two elements, the lower and upper bound of the time to wait before execute
         Returns:
             validity flag: bool, True if the raw trajectory is valid, False if not
             pos_traj: a vector instance of the preprocessed position trajectory 
@@ -97,7 +100,8 @@ class RawInterfaceWrapper(gym.Wrapper):
         """
         return True
 
-    def invalid_traj_callback(self, action: np.ndarray, pos_traj: np.ndarray, vel_traj: np.ndarray) -> Tuple[np.ndarray, float, bool, dict]:
+    def invalid_traj_callback(self, action: np.ndarray, pos_traj: np.ndarray, vel_traj: np.ndarray,
+                              tau_bound: list, delay_bound: list) -> Tuple[np.ndarray, float, bool, dict]:
         """
         Used to return a artificial return from the env if the desired trajectory is invalid.
         Args:
@@ -105,6 +109,8 @@ class RawInterfaceWrapper(gym.Wrapper):
             specified, else only traj_gen parameters
             pos_traj: a vector instance of the raw position trajectory
             vel_traj: a vector instance of the raw velocity trajectory
+            tau_bound: a list of two elements, the lower and upper bound of the trajectory length scaling factor
+            delay_bound: a list of two elements, the lower and upper bound of the time to wait before execute
         Returns:
             obs: artificial observation if the trajectory is invalid, by default a zero vector
             reward: artificial reward if the trajectory is invalid, by default 0

fancy_gym/envs/__init__.py

@@ -237,6 +237,12 @@ for reward_type in ["Dense", "TemporalSparse", "TemporalSpatialSparse"]:
         entry_point='fancy_gym.envs.mujoco:BoxPushing{}'.format(reward_type),
         max_episode_steps=MAX_EPISODE_STEPS_BOX_PUSHING,
     )
+    register(
+        id='BoxPushingRandomInit{}-v0'.format(reward_type),
+        entry_point='fancy_gym.envs.mujoco:BoxPushing{}'.format(reward_type),
+        max_episode_steps=MAX_EPISODE_STEPS_BOX_PUSHING,
+        kwargs={"random_init": True}
+    )
 
 # Here we use the same reward as in BeerPong-v0, but now consider after the release,
 # only one time step, i.e. we simulate until the end of th episode
@@ -500,7 +506,9 @@ for _v in _versions:
 # ########################################################################################################################
 
 ## Box Pushing
-_versions = ['BoxPushingDense-v0', 'BoxPushingTemporalSparse-v0', 'BoxPushingTemporalSpatialSparse-v0']
+_versions = ['BoxPushingDense-v0', 'BoxPushingTemporalSparse-v0', 'BoxPushingTemporalSpatialSparse-v0',
+             'BoxPushingRandomInitDense-v0', 'BoxPushingRandomInitTemporalSparse-v0',
+             'BoxPushingRandomInitTemporalSpatialSparse-v0']
 for _v in _versions:
     _name = _v.split("-")
     _env_id = f'{_name[0]}ProMP-{_name[1]}'
@@ -518,6 +526,27 @@ for _v in _versions:
     )
     ALL_FANCY_MOVEMENT_PRIMITIVE_ENVIRONMENTS["ProMP"].append(_env_id)
 
+for _v in _versions:
+    _name = _v.split("-")
+    _env_id = f'{_name[0]}ProDMP-{_name[1]}'
+    kwargs_dict_box_pushing_prodmp = deepcopy(DEFAULT_BB_DICT_ProDMP)
+    kwargs_dict_box_pushing_prodmp['wrappers'].append(mujoco.box_pushing.MPWrapper)
+    kwargs_dict_box_pushing_prodmp['name'] = _v
+    kwargs_dict_box_pushing_prodmp['controller_kwargs']['p_gains'] = 0.01 * np.array([120., 120., 120., 120., 50., 30., 10.])
+    kwargs_dict_box_pushing_prodmp['controller_kwargs']['d_gains'] = 0.01 * np.array([10., 10., 10., 10., 6., 5., 3.])
+    kwargs_dict_box_pushing_prodmp['trajectory_generator_kwargs']['weights_scale'] = 0.3
+    kwargs_dict_box_pushing_prodmp['trajectory_generator_kwargs']['goal_scale'] = 0.3
+    kwargs_dict_box_pushing_prodmp['trajectory_generator_kwargs']['auto_scale_basis'] = True
+    kwargs_dict_box_pushing_prodmp['basis_generator_kwargs']['num_basis'] = 4
+    kwargs_dict_box_pushing_prodmp['basis_generator_kwargs']['basis_bandwidth_factor'] = 3
+    kwargs_dict_box_pushing_prodmp['phase_generator_kwargs']['alpha_phase'] = 3
+    register(
+        id=_env_id,
+        entry_point='fancy_gym.utils.make_env_helpers:make_bb_env_helper',
+        kwargs=kwargs_dict_box_pushing_prodmp
+    )
+    ALL_FANCY_MOVEMENT_PRIMITIVE_ENVIRONMENTS["ProDMP"].append(_env_id)
+
 for _v in _versions:
     _name = _v.split("-")
     _env_id = f'{_name[0]}ReplanProDMP-{_name[1]}'
@@ -529,9 +558,7 @@ for _v in _versions:
     kwargs_dict_box_pushing_prodmp['trajectory_generator_kwargs']['weights_scale'] = 0.3
     kwargs_dict_box_pushing_prodmp['trajectory_generator_kwargs']['goal_scale'] = 0.3
     kwargs_dict_box_pushing_prodmp['trajectory_generator_kwargs']['auto_scale_basis'] = True
-    kwargs_dict_box_pushing_prodmp['trajectory_generator_kwargs']['goal_offset'] = 1.0
+    kwargs_dict_box_pushing_prodmp['basis_generator_kwargs']['num_basis'] = 4
-    kwargs_dict_box_pushing_prodmp['trajectory_generator_kwargs']['disable_goal'] = True
-    kwargs_dict_box_pushing_prodmp['basis_generator_kwargs']['num_basis'] = 5
     kwargs_dict_box_pushing_prodmp['basis_generator_kwargs']['basis_bandwidth_factor'] = 3
     kwargs_dict_box_pushing_prodmp['phase_generator_kwargs']['alpha_phase'] = 3
     kwargs_dict_box_pushing_prodmp['black_box_kwargs']['max_planning_times'] = 4
@@ -557,8 +584,8 @@ for _v in _versions:
     kwargs_dict_tt_promp['name'] = _v
     kwargs_dict_tt_promp['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_promp['controller_kwargs']['d_gains'] = 0.5 * np.array([0.1, 0.4, 0.2, 0.4, 0.1, 0.4, 0.1])
-    kwargs_dict_tt_promp['phase_generator_kwargs']['learn_tau'] = False
+    kwargs_dict_tt_promp['phase_generator_kwargs']['learn_tau'] = True
-    kwargs_dict_tt_promp['phase_generator_kwargs']['learn_delay'] = False
+    kwargs_dict_tt_promp['phase_generator_kwargs']['learn_delay'] = True
     kwargs_dict_tt_promp['phase_generator_kwargs']['tau_bound'] = [0.8, 1.5]
     kwargs_dict_tt_promp['phase_generator_kwargs']['delay_bound'] = [0.05, 0.15]
     kwargs_dict_tt_promp['basis_generator_kwargs']['num_basis'] = 3

fancy_gym/envs/mujoco/box_pushing/box_pushing_env.py

@@ -370,8 +370,8 @@ class BoxPushingTemporalSpatialSparse(BoxPushingEnvBase):
 
 class BoxPushingTemporalSpatialSparse2(BoxPushingEnvBase):
 
-    def __init__(self, frame_skip: int = 10):
+    def __init__(self, frame_skip: int = 10, random_init: bool = False):
-        super(BoxPushingTemporalSpatialSparse2, self).__init__(frame_skip=frame_skip)
+        super(BoxPushingTemporalSpatialSparse2, self).__init__(frame_skip=frame_skip, random_init=random_init)
 
     def _get_reward(self, episode_end, box_pos, box_quat, target_pos, target_quat,
                     rod_tip_pos, rod_quat, qpos, qvel, action):
@@ -405,9 +405,9 @@ class BoxPushingTemporalSpatialSparse2(BoxPushingEnvBase):
         return reward
 
 
-class BoxPushingBruceSparse(BoxPushingEnvBase):
+class BoxPushingNoConstraintSparse(BoxPushingEnvBase):
-    def __init__(self, frame_skip: int = 10):
+    def __init__(self, frame_skip: int = 10, random_init: bool = False):
-        super(BoxPushingBruceSparse, self).__init__(frame_skip=frame_skip)
+        super(BoxPushingNoConstraintSparse, self).__init__(frame_skip=frame_skip, random_init=random_init)
 
     def _get_reward(self, episode_end, box_pos, box_quat, target_pos, target_quat,
                     rod_tip_pos, rod_quat, qpos, qvel, action):
@@ -435,16 +435,3 @@ class BoxPushingBruceSparse(BoxPushingEnvBase):
         box_rot_vel = box_rot_pos_vel[:3]
         box_pos_vel = box_rot_pos_vel[3:]
         return -rot_coeff * np.linalg.norm(box_rot_vel) - pos_coeff * np.linalg.norm(box_pos_vel)
-
-
-
-if __name__=="__main__":
-    import fancy_gym
-    env = fancy_gym.make("BoxPushingDenseProDMP-v0", seed=0)
-    env.reset()
-    for i in range(1000):
-        env.render(mode="human")
-        obs, rew, done, info = env.step(env.action_space.sample())
-        if done:
-            env.reset()
-            # print(f"box_end_velocity: {info['box_end_vel']}")

fancy_gym/envs/mujoco/table_tennis/mp_wrapper.py

@@ -29,15 +29,16 @@ class TT_MPWrapper(RawInterfaceWrapper):
     def current_vel(self) -> Union[float, int, np.ndarray, Tuple]:
         return self.data.qvel[:7].copy()
 
-    def preprocessing_and_validity_callback(self, action, pos_traj, vel_traj):
+    def preprocessing_and_validity_callback(self, action: np.ndarray, pos_traj: np.ndarray, vel_traj: np.ndarray,
-        return self.check_traj_validity(action, pos_traj, vel_traj)
+                                            tau_bound: list, delay_bound:list):
+        return self.check_traj_validity(action, pos_traj, vel_traj, tau_bound, delay_bound)
 
     def set_episode_arguments(self, action, pos_traj, vel_traj):
         return pos_traj, vel_traj
 
     def invalid_traj_callback(self, action: np.ndarray, pos_traj: np.ndarray, vel_traj: np.ndarray,
-                              return_contextual_obs: bool) -> Tuple[np.ndarray, float, bool, dict]:
+                              return_contextual_obs: bool, tau_bound:list, delay_bound:list) -> Tuple[np.ndarray, float, bool, dict]:
-        return self.get_invalid_traj_step_return(action, pos_traj, return_contextual_obs)
+        return self.get_invalid_traj_step_return(action, pos_traj, return_contextual_obs, tau_bound, delay_bound)
 
 class TTVelObs_MPWrapper(TT_MPWrapper):
 

fancy_gym/envs/mujoco/table_tennis/table_tennis_env.py

@@ -5,7 +5,7 @@ from gym import utils, spaces
 from gym.envs.mujoco import MujocoEnv
 
 from fancy_gym.envs.mujoco.table_tennis.table_tennis_utils import is_init_state_valid, magnus_force
-from fancy_gym.envs.mujoco.table_tennis.table_tennis_utils import jnt_pos_low, jnt_pos_high, delay_bound, tau_bound
+from fancy_gym.envs.mujoco.table_tennis.table_tennis_utils import jnt_pos_low, jnt_pos_high
 
 import mujoco
 
@@ -225,7 +225,7 @@ class TableTennisEnv(MujocoEnv, utils.EzPickle):
             init_ball_state = self._generate_random_ball(random_pos=random_pos, random_vel=random_vel)
         return init_ball_state
 
-    def _get_traj_invalid_penalty(self, action, pos_traj):
+    def _get_traj_invalid_penalty(self, action, pos_traj, tau_bound, delay_bound):
         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]]))
         violate_high_bound_error = np.mean(np.maximum(pos_traj - jnt_pos_high, 0))
@@ -234,9 +234,9 @@ class TableTennisEnv(MujocoEnv, utils.EzPickle):
                           violate_high_bound_error + violate_low_bound_error
         return -invalid_penalty
 
-    def get_invalid_traj_step_return(self, action, pos_traj, contextual_obs):
+    def get_invalid_traj_step_return(self, action, pos_traj, contextual_obs, tau_bound, delay_bound):
         obs = self._get_obs() if contextual_obs else np.concatenate([self._get_obs(), np.array([0])]) # 0 for invalid traj
-        penalty = self._get_traj_invalid_penalty(action, pos_traj)
+        penalty = self._get_traj_invalid_penalty(action, pos_traj, tau_bound, delay_bound)
         return obs, penalty, True, {
             "hit_ball": [False],
             "ball_returned_success": [False],
@@ -247,7 +247,7 @@ class TableTennisEnv(MujocoEnv, utils.EzPickle):
         }
 
     @staticmethod
-    def check_traj_validity(action, pos_traj, vel_traj):
+    def check_traj_validity(action, pos_traj, vel_traj, tau_bound, delay_bound):
         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):