small bp and tt updates

alr_envs/alr/mujoco/beerpong/beerpong.py

@@ -127,24 +127,28 @@ class ALRBeerBongEnv(MujocoEnv, utils.EzPickle):
             self._steps += 1
         else:
             reward = -30
+            reward_infos = dict()
             success = False
             is_collided = False
             done = True
             ball_pos = np.zeros(3)
             ball_vel = np.zeros(3)
 
-        return ob, reward, done, dict(reward_dist=reward_dist,
-                                      reward_ctrl=reward_ctrl,
-                                      reward=reward,
-                                      velocity=angular_vel,
-                                      # traj=self._q_pos,
-                                      action=a,
-                                      q_pos=self.sim.data.qpos[0:7].ravel().copy(),
-                                      q_vel=self.sim.data.qvel[0:7].ravel().copy(),
-                                      ball_pos=ball_pos,
-                                      ball_vel=ball_vel,
-                                      success=success,
-                                      is_collided=is_collided, sim_crash=crash)
+        infos = dict(reward_dist=reward_dist,
+                     reward_ctrl=reward_ctrl,
+                     reward=reward,
+                     velocity=angular_vel,
+                     # traj=self._q_pos,
+                     action=a,
+                     q_pos=self.sim.data.qpos[0:7].ravel().copy(),
+                     q_vel=self.sim.data.qvel[0:7].ravel().copy(),
+                     ball_pos=ball_pos,
+                     ball_vel=ball_vel,
+                     success=success,
+                     is_collided=is_collided, sim_crash=crash)
+        infos.update(reward_infos)
+
+        return ob, reward, done, infos
 
     def check_traj_in_joint_limits(self):
         return any(self.current_pos > self.j_max) or any(self.current_pos < self.j_min)

alr_envs/alr/mujoco/beerpong/beerpong_reward_staged.py

@@ -110,7 +110,7 @@ class BeerPongReward:
             success = ball_in_cup
             crash = self._is_collided
         else:
-            reward = - 1e-4 * action_cost
+            reward = - 1e-2 * action_cost
             success = False
             crash = False
 

alr_envs/alr/mujoco/table_tennis/tt_gym.py

@@ -10,7 +10,7 @@ from alr_envs.alr.mujoco.table_tennis.tt_reward import TT_Reward
 
 #TODO: Check for simulation stability. Make sure the code runs even for sim crash
 
-MAX_EPISODE_STEPS = 2875
+MAX_EPISODE_STEPS = 1750
 BALL_NAME_CONTACT = "target_ball_contact"
 BALL_NAME = "target_ball"
 TABLE_NAME = 'table_tennis_table'
@@ -42,9 +42,10 @@ class TTEnvGym(MujocoEnv, utils.EzPickle):
         else:
             raise ValueError("either 2 or 4 dimensional Contexts available")
 
-        action_space_low = np.array([-2.6, -2.0, -2.8, -0.9, -4.8, -1.6, -2.2])
-        action_space_high = np.array([2.6, 2.0, 2.8, 3.1, 1.3, 1.6, 2.2])
-        self.action_space = spaces.Box(low=action_space_low, high=action_space_high, dtype='float64')
+        # has no effect as it is overwritten in init of super
+        # action_space_low = np.array([-2.6, -2.0, -2.8, -0.9, -4.8, -1.6, -2.2])
+        # action_space_high = np.array([2.6, 2.0, 2.8, 3.1, 1.3, 1.6, 2.2])
+        # self.action_space = spaces.Box(low=action_space_low, high=action_space_high, dtype='float64')
 
         self.time_steps = 0
         self.init_qpos_tt = np.array([0, 0, 0, 1.5, 0, 0, 1.5, 0, 0, 0])
@@ -159,7 +160,10 @@ class TTEnvGym(MujocoEnv, utils.EzPickle):
             done = True
             reward = -25
         ob = self._get_obs()
-        return ob, reward, done, {"hit_ball": self.hit_ball}  # might add some information here ....
+        info = {"hit_ball": self.hit_ball,
+                "q_pos": np.copy(self.sim.data.qpos[:7]),
+                "ball_pos": np.copy(self.sim.data.qpos[7:])}
+        return ob, reward, done, info # might add some information here ....
 
     def set_context(self, context):
         old_state = self.sim.get_state()

alr_envs/alr/mujoco/table_tennis/tt_reward.py

@@ -19,7 +19,7 @@ class TT_Reward:
             # # seems to work for episodic case
             min_r_b_dist = np.min(np.linalg.norm(np.array(self.c_ball_traj) - np.array(self.c_racket_traj), axis=1))
             if not hited_ball:
-                return 0.2 * (1- np.tanh(min_r_b_dist**2))
+                return 0.2 * (1 - np.tanh(min_r_b_dist**2))
             else:
                 if ball_landing_pos is None:
                     min_b_des_b_dist = np.min(np.linalg.norm(np.array(self.c_ball_traj)[:,:2] - self.c_goal[:2], axis=1))