Dev

README.md

@@ -73,7 +73,7 @@ MultiTask Suite
    This suite contains a collection of environments centered around dexterous manipulation. Standard [TCDM benchmarks](https://pregrasps.github.io/) are a part of this suite
 
 ## - ROBEL Suite (Coming soon)
-   This suite contains a collection of environments centered around real-world locomotion and manipulation. Standard [ROBEL benchmarks](http://roboticsbenchmarks.org/) are a part of this suite
+   This suite contains a collection of environments centered around real-world locomotion and manipulation. Standard [ROBEL benchmarks](https://sites.google.com/view/roboticsbenchmarks) are a part of this suite
 
 # Citation
 If you find `RoboHive` useful in your research,

robohive/envs/env_base.py

@@ -495,8 +495,8 @@ def get_input_seed(self):
 
     def _reset(self, reset_qpos=None, reset_qvel=None, seed=None, **kwargs):
         """
-        Reset the environment
-        Default implemention provided. Override if env needs custom reset
+        Reset the environment (Default implemention provided).
+        Override if env needs custom reset. Carefully handle return type for gym/gymnasium compatibility
         """
         qpos = self.init_qpos.copy() if reset_qpos is None else reset_qpos
         qvel = self.init_qvel.copy() if reset_qvel is None else reset_qvel

robohive/envs/hands/baoding_v1.py

@@ -264,8 +264,7 @@ def reset(self, reset_pose=None, reset_vel=None, reset_goal=None, time_period=6,
         self.goal = self.create_goal_trajectory(time_period=time_period) if reset_goal is None else reset_goal.copy()
 
         # reset scene
-        obs = super().reset(reset_qpos=reset_pose, reset_qvel=reset_vel, **kwargs)
-        return obs
+        return super().reset(reset_qpos=reset_pose, reset_qvel=reset_vel, **kwargs)
 
     def create_goal_trajectory(self, time_step=.1, time_period=6):
         len_of_goals = 1000 # assumes that its greator than env horizon
@@ -326,5 +325,4 @@ def create_goal_trajectory(self, time_step=.1, time_period=6):
 class BaodingRandomEnvV1(BaodingFixedEnvV1):
 
     def reset(self, **kwargs):
-        obs = super().reset(time_period = self.np_random.uniform(high=5, low=7), **kwargs)
-        return obs
+        return super().reset(time_period = self.np_random.uniform(high=5, low=7), **kwargs)

robohive/envs/hands/door_v1.py

@@ -96,17 +96,13 @@ def get_reward_dict(self, obs_dict):
         return rwd_dict
 
 
-    def reset(self, reset_qpos=None, reset_qvel=None, **kwargs):
+    def reset(self, **kwargs):
         self.sim.reset()
-        qp = self.init_qpos.copy() if reset_qpos is None else reset_qpos
-        qv = self.init_qvel.copy() if reset_qvel is None else reset_qvel
-        self.robot.reset(reset_pos=qp, reset_vel=qv, **kwargs)
-
         self.sim.model.body_pos[self.door_bid,0] = self.np_random.uniform(low=-0.3, high=-0.2)
         self.sim.model.body_pos[self.door_bid, 1] = self.np_random.uniform(low=0.25, high=0.35)
         self.sim.model.body_pos[self.door_bid,2] = self.np_random.uniform(low=0.252, high=0.35)
         self.sim.forward()
-        return self.get_obs()
+        return super().reset(**kwargs)
 
 
     def get_env_state(self):

robohive/envs/hands/hammer_v1.py

@@ -120,15 +120,12 @@ def get_obs_dict(self, sim):
         return obs_dict
 
 
-    def reset(self, reset_qpos=None, reset_qvel=None, **kwargs):
+    def reset(self, **kwargs):
         self.sim.reset()
-        qp = self.init_qpos.copy() if reset_qpos is None else reset_qpos
-        qv = self.init_qvel.copy() if reset_qvel is None else reset_qvel
-        self.robot.reset(reset_pos=qp, reset_vel=qv, **kwargs)
-
         self.sim.model.body_pos[self.target_bid,2] = self.np_random.uniform(low=0.1, high=0.25)
         self.sim.forward()
-        return self.get_obs()
+        return super().reset(**kwargs)
+
 
     def get_env_state(self):
         """

robohive/envs/hands/pen_v1.py

@@ -108,19 +108,14 @@ def get_reward_dict(self, obs_dict):
         return rwd_dict
 
 
-    def reset(self, reset_qpos=None, reset_qvel=None, **kwargs):
+    def reset(self, **kwargs):
         self.sim.reset()
-        qp = self.init_qpos.copy() if reset_qpos is None else reset_qpos
-        qv = self.init_qvel.copy() if reset_qvel is None else reset_qvel
-        self.robot.reset(reset_pos=qp, reset_vel=qv, **kwargs)
-
         desired_orien = np.zeros(3)
         desired_orien[0] = self.np_random.uniform(low=-1, high=1)
         desired_orien[1] = self.np_random.uniform(low=-1, high=1)
         self.sim.model.body_quat[self.target_obj_bid] = euler2quat(desired_orien)
         self.sim.forward()
-
-        return self.get_obs()
+        return super().reset(**kwargs)
 
 
     def get_env_state(self):

robohive/envs/hands/relocate_v1.py

@@ -136,20 +136,16 @@ def get_obs_dict(self, sim):
         return obs_dict
 
 
-    def reset(self, reset_qpos=None, reset_qvel=None, **kwargs):
+    def reset(self, **kwargs):
         self.sim.reset()
-        qp = self.init_qpos.copy() if reset_qpos is None else reset_qpos
-        qv = self.init_qvel.copy() if reset_qvel is None else reset_qvel
-        self.robot.reset(reset_pos=qp, reset_vel=qv, **kwargs)
-
-
         self.sim.model.body_pos[self.obj_bid,0] = self.np_random.uniform(low=-0.15, high=0.15)
         self.sim.model.body_pos[self.obj_bid,1] = self.np_random.uniform(low=-0.15, high=0.3)
         self.sim.model.site_pos[self.target_obj_sid, 0] = self.np_random.uniform(low=-0.2, high=0.2)
         self.sim.model.site_pos[self.target_obj_sid,1] = self.np_random.uniform(low=-0.2, high=0.2)
         self.sim.model.site_pos[self.target_obj_sid,2] = self.np_random.uniform(low=0.15, high=0.35)
         self.sim.forward()
-        return self.get_obs()
+        return super().reset(**kwargs)
+
 
     def get_env_state(self):
         """

robohive/envs/multi_task/README.md

@@ -2,7 +2,7 @@
 This suite is designed to study generalization in multi-task settings. RoboHive's multi-task suite builds from `FrankaKitchen` environements originally studied in the [Relay Policy Learning](https://relay-policy-learning.github.io/) project.
 
 # Franka Kitchen
-`FrankaKitchen` domain offers a challenging set of manipulation problems in an unstructured environment with many possible tasks to perform. The original set consisted of a franka robot in a kitchen domain. Overtime, Franka Kitchen has became a popular choice of environments for studying multi-task generalization. Its widespread use has led to a few different publically available variations. To help navigate these changes, we name these variations and document its evolution of across various versions below -
+`FrankaKitchen` domain offers a challenging set of manipulation problems in an unstructured environment with many possible tasks to perform. The original set consisted of a franka robot in a kitchen domain. Overtime, Franka Kitchen has became a popular choice of environments for studying multi-task generalization. Its widespread use has led to a few different publically available variations. To help navigate these changes, we name these variations and document its evolution across various versions below -
 
 ## Change log/ History
 

robohive/envs/multi_task/multi_task_base_v1.py

@@ -200,9 +200,9 @@ def get_reward_dict(self, obs_dict):
                 # Optional Keys
                 ("obj_goal", -np.sum(goal_dist, axis=-1)),
                 ("bonus",
-                    1.0*np.product(goal_dist < 5 * self.obj["dof_proximity"], axis=-1)
+                    1.0*np.prod(goal_dist < 5 * self.obj["dof_proximity"], axis=-1)
                     # np.product(goal_dist < 0.75 * self.obj["dof_ranges"], axis=-1)
-                    + 1.0*np.product(goal_dist < 1.67 * self.obj["dof_proximity"], axis=-1),
+                    + 1.0*np.prod(goal_dist < 1.67 * self.obj["dof_proximity"], axis=-1),
                     # + np.product(goal_dist < 0.25 * self.obj["dof_ranges"], axis=-1),
                 ),
                 ("pose", -np.sum(np.abs(obs_dict["pose_err"]), axis=-1)),
@@ -272,4 +272,4 @@ def set_obj_goal(self, obj_goal=None, interact_site=None):
         elif type(interact_site) is str:  # overwrite using name
             self.interact_sid = self.sim.model.site_name2id(interact_site)
         elif type(interact_site) is int:  # overwrite using id
-            self.interact_sid = interact_site
+            self.interact_sid = interact_site