Dev

.github/workflows/python-app.yml

@@ -23,21 +23,25 @@ jobs:
 
     steps:
     - uses: actions/checkout@v3
+
     - name: Set up Python 3.10
       uses: actions/setup-python@v3
       with:
         python-version: "3.10"
+
     - name: Install mesa and patchelf
       run: |
+        sudo apt-get update -y -qq
         sudo apt-get install -y \
             libgl1-mesa-dev \
-            libgl1-mesa-glx \
+            libgl1 \
+            libglx-mesa0 \
             libglew-dev \
             libosmesa6-dev \
-            software-properties-common
-        sudo apt-get install -y patchelf
-        sudo apt-get update -y -qq
-        sudo apt-get install -y xvfb x11-xserver-utils
+            software-properties-common \
+            patchelf \
+            xvfb \
+            x11-xserver-utils
         sudo /usr/bin/Xvfb $DISPLAY -screen 0 1280x1024x24 &
 
     - name: Install ffmpeg
@@ -46,6 +50,7 @@ jobs:
 
     - name: Checkout submodules
       run: git submodule update --init --recursive
+
     - name: Install dependencies
       run: |
         python -m pip install --upgrade pip
@@ -56,10 +61,6 @@ jobs:
     - name: Run MuJoCo Rendering test
       run: python3 -m mujoco.render_test
 
-    # - name: Test RoboHive (Fetching SimHive via git-submodules)
-    #   run: |
-    #     python robohive/tests/test_all.py
-
     - name: Test RoboHive (Programmatically fetching SimHive)
       run: |
         rm -rf robohive/simhive

.gitmodules

@@ -38,4 +38,4 @@
 	url = https://github.com/vikashplus/robotiq_sim.git
 [submodule "robohive/simhive/myo_sim"]
 	path = robohive/simhive/myo_sim
-	url = https://github.com/MyoHub/myo_sim.git
+	url = https://github.com/MyoHub/myo_sim.git

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

@@ -5,21 +5,21 @@
 License :: Under Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.
 ================================================= """
 
-from robohive.utils import gym
-import numpy as np
 import os
 import time as timer
+from sys import platform
 
-from robohive.envs.obs_vec_dict import ObsVecDict
-from robohive.utils import tensor_utils
-from robohive.robot.robot import Robot
-from robohive.utils.implement_for import implement_for
-from robohive.utils.prompt_utils import prompt, Prompt
+import numpy as np
 import skvideo.io
-from sys import platform
-from robohive.physics.sim_scene import SimScene
+
 import robohive.utils.import_utils as import_utils
 from robohive.envs.env_variants import gym_registry_specs
+from robohive.envs.obs_vec_dict import ObsVecDict
+from robohive.physics.sim_scene import SimScene
+from robohive.robot.robot import Robot
+from robohive.utils import gym, tensor_utils
+from robohive.utils.implement_for import implement_for
+from robohive.utils.prompt_utils import Prompt, prompt
 
 # TODO
 # remove rwd_mode
@@ -183,7 +183,14 @@ def forward(self, x):
             if "rrl" in id_encoder or "resnet" in id_encoder:
                 import_utils.torchvision_isavailable()
                 import torchvision.transforms as T
-                from torchvision.models import resnet50, ResNet50_Weights, resnet34, ResNet34_Weights, resnet18, ResNet18_Weights
+                from torchvision.models import (
+                    ResNet18_Weights,
+                    ResNet34_Weights,
+                    ResNet50_Weights,
+                    resnet18,
+                    resnet34,
+                    resnet50,
+                )
 
             if "r3m" in id_encoder:
                 import_utils.torchvision_isavailable()
@@ -495,8 +502,8 @@ 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

robohive/envs/multi_task/utils/parse_demos.py

@@ -84,7 +84,7 @@ def render_demos(env, data, filename='demo_rendering.mp4', render=None):
 
 # playback demos and get data(physics respected)
 def gather_training_data(env, data, filename='demo_playback.mp4', render=None):
-    env = env.env
+    env = env.unwrapped
     FPS = 30
     render_skip = max(1, round(1. / \
         (FPS * env.sim.model.opt.timestep * env.frame_skip)))
@@ -127,7 +127,7 @@ def gather_training_data(env, data, filename='demo_playback.mp4', render=None):
         # normalization and env stepping
         act = (ctrl - act_mid) / act_rng
         act = np.clip(act, -0.999, 0.999)
-        next_obs, reward, done, env_info = env.step(act)
+        next_obs, reward, done, *_, env_info = env.step(act)
         path_reward += reward
 
         # populate path

robohive/envs/myo/base_v0.py

@@ -4,6 +4,7 @@
 ================================================= """
 
 from robohive.envs import env_base
+import mujoco
 import numpy as np
 
 class BaseV0(env_base.MujocoEnv):
@@ -88,15 +89,24 @@ def step(self, a, **kwargs):
 
         # implement abnormalities
         if self.muscle_condition == 'fatigue':
+            actuator_moment = np.zeros((self.sim.model.nu, self.sim.model.nv))
+            mujoco.mju_sparse2dense(
+                actuator_moment,
+                self.sim.data.actuator_moment.reshape(-1),
+                self.sim.data.moment_rownnz,
+                self.sim.data.moment_rowadr,
+                self.sim.data.moment_colind.reshape(-1),
+            )
             for mus_idx in range(self.sim.model.actuator_gainprm.shape[0]):
 
-                if self.sim.data.actuator_moment.shape[1]==1:
-                    self.f_load[mus_idx].append(self.sim.data.actuator_moment[mus_idx].copy())
+                if actuator_moment.shape[1] == 1:
+                    self.f_load[mus_idx].append(actuator_moment[mus_idx].copy())
                 else:
-                    self.f_load[mus_idx].append(self.sim.data.actuator_moment[mus_idx,1].copy())
+                    self.f_load[mus_idx].append(actuator_moment[mus_idx, 1].copy())
 
                 if self.MVC_rest[mus_idx] != 0:
-                    f_int = np.sum(self.f_load[mus_idx]-np.max(self.f_load[mus_idx],0),0)/self.MVC_rest[mus_idx]
+                    f_load = np.asarray(self.f_load[mus_idx])
+                    f_int = np.sum(f_load - np.max(f_load, 0), 0)/self.MVC_rest[mus_idx]
                     f_cem = self.MVC_rest[mus_idx]*np.exp(self.k_fatigue*f_int)
                 else:
                     f_cem = 0

robohive/envs/myo/myodm/myodm_v0.py

@@ -285,7 +285,7 @@ def playback(self):
         ref_mot = self.ref.get_reference(self.time+self.motion_start_time)
         self.qpos_from_robot_object(self.sim.data.qpos, ref_mot.robot, ref_mot.object )
         self.sim.forward()
-        self.sim.data.time = self.sim.data.time + 0.02#self.env.env.dt
+        self.sim.data.time = self.sim.data.time + 0.02#self.env.unwrapped.dt
         return idxs[0] < self.ref.horizon-1
 
 

robohive/envs/obs_vec_dict.py

@@ -5,9 +5,11 @@
 License :: Under Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.
 ================================================= """
 
-import numpy as np
 from collections import deque
 
+import numpy as np
+
+
 class ObsVecDict():
     """
     Class to help with conversion between obs_dict <> obs_vector
@@ -79,7 +81,7 @@ def obsdict2obsvec(self, obs_dict, ordered_obs_keys):
             self.initialize(obs_dict, ordered_obs_keys)
 
         # recover vec
-        obsvec = np.zeros(0)
+        obsvec = np.zeros(0, dtype=np.float32)
         for key in self.ordered_obs_keys:
             obsvec = np.concatenate([obsvec, obs_dict[key].ravel()]) # ravel helps with images
 

robohive/envs/tcdm/playback_mocap.py

@@ -23,9 +23,9 @@
     args = parser.parse_args()
     args.end = args.end if args.end is not None else args.start + args.length
 
-    # envMyoSuit=gym.make('MyoHandAirplanePass-v0')
-    envMyoSuit=gym.make(args.sim_name)
-    physics_myo=envMyoSuit.env.sim
+    # envMyoSuite=gym.make('MyoHandAirplanePass-v0')
+    envMyoSuite=gym.make(args.sim_name)
+    physics_myo=envMyoSuite.unwrapped.sim
     # jx_mocap coordinates
     # for i,b in enumerate(BODIES): #print(b,i, physics_myo.sim.named.data.xipos[b])
     #     print(f"<body mocap=\"true\" name=\"j{i}_mocap\" pos=\"{' '.join(map(str, physics_myo.sim.named.data.xipos[b]))} \" ><site name=\"j{i}\" size=\"0.015\" rgba=\"0 0 1 0.5\" pos=\"0 0 0\"/></body>\" ")

robohive/envs/tcdm/track.py

@@ -273,7 +273,7 @@ def playback(self):
         ref_mot = self.ref.get_reference(self.time)
         self.qpos_from_robot_object(self.sim.data.qpos, ref_mot.robot, ref_mot.object )
         self.sim.forward()
-        self.sim.data.time = self.sim.data.time + 0.02#self.env.env.dt
+        self.sim.data.time = self.sim.data.time + 0.02#self.env.unwrapped.dt
         return idxs[0] < self.ref.horizon-1