Migrate to mujoco-py 1.50 (openai#834)

* all envs run offscreen

* render works

* changed mujoco-py version

* Bump versions

* Update version and README

* Same versioning for all mujoco envs

* Fix typo

* Fix version

* Bump version again

* Revert "Fix version"

This reverts commit decc577.

README.rst

@@ -4,7 +4,7 @@ OpenAI Gym
 **OpenAI Gym is a toolkit for developing and comparing reinforcement learning algorithms.** This is the ``gym`` open-source library, which gives you access to a standardized set of environments.
 
 .. image:: https://travis-ci.org/openai/gym.svg?branch=master
-    :target: https://travis-ci.org/openai/gym
+	:target: https://travis-ci.org/openai/gym
 
 `See What's New section below <#what-s-new>`_
 
@@ -126,7 +126,7 @@ fake display. The easiest way to do this is by running under
 
 .. code:: shell
 
-     xvfb-run -s "-screen 0 1400x900x24" bash
+	 xvfb-run -s "-screen 0 1400x900x24" bash
 
 Installing dependencies for specific environments
 -------------------------------------------------
@@ -261,6 +261,10 @@ We are using `pytest <http://doc.pytest.org>`_ for tests. You can run them via:
 What's new
 ==========
 
+- 2018-01-24: All continuous control environments now use mujoco_py >= 1.50.
+  Versions have been updated accordingly to -v2, e.g. HalfCheetah-v2. Performance
+  should be similar (see https://github.com/openai/gym/pull/834) but there are likely
+  some differences due to changes in MuJoCo.
 - 2017-06-16: Make env.spec into a property to fix a bug that occurs
   when you try to print out an unregistered Env.
 - 2017-05-13: BACKWARDS INCOMPATIBILITY: The Atari environments are now at

gym/envs/__init__.py

@@ -204,89 +204,89 @@
 # 2D
 
 register(
-    id='Reacher-v1',
+    id='Reacher-v2',
     entry_point='gym.envs.mujoco:ReacherEnv',
     max_episode_steps=50,
     reward_threshold=-3.75,
 )
 
 register(
-    id='Pusher-v0',
+    id='Pusher-v2',
     entry_point='gym.envs.mujoco:PusherEnv',
     max_episode_steps=100,
     reward_threshold=0.0,
 )
 
 register(
-    id='Thrower-v0',
+    id='Thrower-v2',
     entry_point='gym.envs.mujoco:ThrowerEnv',
     max_episode_steps=100,
     reward_threshold=0.0,
 )
 
 register(
-    id='Striker-v0',
+    id='Striker-v2',
     entry_point='gym.envs.mujoco:StrikerEnv',
     max_episode_steps=100,
     reward_threshold=0.0,
 )
 
 register(
-    id='InvertedPendulum-v1',
+    id='InvertedPendulum-v2',
     entry_point='gym.envs.mujoco:InvertedPendulumEnv',
     max_episode_steps=1000,
     reward_threshold=950.0,
 )
 
 register(
-    id='InvertedDoublePendulum-v1',
+    id='InvertedDoublePendulum-v2',
     entry_point='gym.envs.mujoco:InvertedDoublePendulumEnv',
     max_episode_steps=1000,
     reward_threshold=9100.0,
 )
 
 register(
-    id='HalfCheetah-v1',
+    id='HalfCheetah-v2',
     entry_point='gym.envs.mujoco:HalfCheetahEnv',
     max_episode_steps=1000,
     reward_threshold=4800.0,
 )
 
 register(
-    id='Hopper-v1',
+    id='Hopper-v2',
     entry_point='gym.envs.mujoco:HopperEnv',
     max_episode_steps=1000,
     reward_threshold=3800.0,
 )
 
 register(
-    id='Swimmer-v1',
+    id='Swimmer-v2',
     entry_point='gym.envs.mujoco:SwimmerEnv',
     max_episode_steps=1000,
     reward_threshold=360.0,
 )
 
 register(
-    id='Walker2d-v1',
+    id='Walker2d-v2',
     max_episode_steps=1000,
     entry_point='gym.envs.mujoco:Walker2dEnv',
 )
 
 register(
-    id='Ant-v1',
+    id='Ant-v2',
     entry_point='gym.envs.mujoco:AntEnv',
     max_episode_steps=1000,
     reward_threshold=6000.0,
 )
 
 register(
-    id='Humanoid-v1',
+    id='Humanoid-v2',
     entry_point='gym.envs.mujoco:HumanoidEnv',
     max_episode_steps=1000,
 )
 
 register(
-    id='HumanoidStandup-v1',
+    id='HumanoidStandup-v2',
     entry_point='gym.envs.mujoco:HumanoidStandupEnv',
     max_episode_steps=1000,
 )

gym/envs/mujoco/ant.py

@@ -14,7 +14,7 @@ def _step(self, a):
         forward_reward = (xposafter - xposbefore)/self.dt
         ctrl_cost = .5 * np.square(a).sum()
         contact_cost = 0.5 * 1e-3 * np.sum(
-            np.square(np.clip(self.model.data.cfrc_ext, -1, 1)))
+            np.square(np.clip(self.sim.data.cfrc_ext, -1, 1)))
         survive_reward = 1.0
         reward = forward_reward - ctrl_cost - contact_cost + survive_reward
         state = self.state_vector()
@@ -30,9 +30,9 @@ def _step(self, a):
 
     def _get_obs(self):
         return np.concatenate([
-            self.model.data.qpos.flat[2:],
-            self.model.data.qvel.flat,
-            np.clip(self.model.data.cfrc_ext, -1, 1).flat,
+            self.sim.data.qpos.flat[2:],
+            self.sim.data.qvel.flat,
+            np.clip(self.sim.data.cfrc_ext, -1, 1).flat,
         ])
 
     def reset_model(self):

gym/envs/mujoco/half_cheetah.py

@@ -8,9 +8,9 @@ def __init__(self):
         utils.EzPickle.__init__(self)
 
     def _step(self, action):
-        xposbefore = self.model.data.qpos[0, 0]
+        xposbefore = self.sim.data.qpos[0]
         self.do_simulation(action, self.frame_skip)
-        xposafter = self.model.data.qpos[0, 0]
+        xposafter = self.sim.data.qpos[0]
         ob = self._get_obs()
         reward_ctrl = - 0.1 * np.square(action).sum()
         reward_run = (xposafter - xposbefore)/self.dt
@@ -20,8 +20,8 @@ def _step(self, action):
 
     def _get_obs(self):
         return np.concatenate([
-            self.model.data.qpos.flat[1:],
-            self.model.data.qvel.flat,
+            self.sim.data.qpos.flat[1:],
+            self.sim.data.qvel.flat,
         ])
 
     def reset_model(self):

gym/envs/mujoco/hopper.py

@@ -8,9 +8,9 @@ def __init__(self):
         utils.EzPickle.__init__(self)
 
     def _step(self, a):
-        posbefore = self.model.data.qpos[0, 0]
+        posbefore = self.sim.data.qpos[0]
         self.do_simulation(a, self.frame_skip)
-        posafter, height, ang = self.model.data.qpos[0:3, 0]
+        posafter, height, ang = self.sim.data.qpos[0:3]
         alive_bonus = 1.0
         reward = (posafter - posbefore) / self.dt
         reward += alive_bonus
@@ -23,8 +23,8 @@ def _step(self, a):
 
     def _get_obs(self):
         return np.concatenate([
-            self.model.data.qpos.flat[1:],
-            np.clip(self.model.data.qvel.flat, -10, 10)
+            self.sim.data.qpos.flat[1:],
+            np.clip(self.sim.data.qvel.flat, -10, 10)
         ])
 
     def reset_model(self):

gym/envs/mujoco/humanoid.py

@@ -2,9 +2,9 @@
 from gym.envs.mujoco import mujoco_env
 from gym import utils
 
-def mass_center(model):
-    mass = model.body_mass
-    xpos = model.data.xipos
+def mass_center(model, sim):
+    mass = np.expand_dims(model.body_mass, 1)
+    xpos = sim.data.xipos
     return (np.sum(mass * xpos, 0) / np.sum(mass))[0]
 
 class HumanoidEnv(mujoco_env.MujocoEnv, utils.EzPickle):
@@ -13,7 +13,7 @@ def __init__(self):
         utils.EzPickle.__init__(self)
 
     def _get_obs(self):
-        data = self.model.data
+        data = self.sim.data
         return np.concatenate([data.qpos.flat[2:],
                                data.qvel.flat,
                                data.cinert.flat,
@@ -22,17 +22,17 @@ def _get_obs(self):
                                data.cfrc_ext.flat])
 
     def _step(self, a):
-        pos_before = mass_center(self.model)
+        pos_before = mass_center(self.model, self.sim)
         self.do_simulation(a, self.frame_skip)
-        pos_after = mass_center(self.model)
+        pos_after = mass_center(self.model, self.sim)
         alive_bonus = 5.0
-        data = self.model.data
+        data = self.sim.data
         lin_vel_cost = 0.25 * (pos_after - pos_before) / self.model.opt.timestep
         quad_ctrl_cost = 0.1 * np.square(data.ctrl).sum()
         quad_impact_cost = .5e-6 * np.square(data.cfrc_ext).sum()
         quad_impact_cost = min(quad_impact_cost, 10)
         reward = lin_vel_cost - quad_ctrl_cost - quad_impact_cost + alive_bonus
-        qpos = self.model.data.qpos
+        qpos = self.sim.data.qpos
         done = bool((qpos[2] < 1.0) or (qpos[2] > 2.0))
         return self._get_obs(), reward, done, dict(reward_linvel=lin_vel_cost, reward_quadctrl=-quad_ctrl_cost, reward_alive=alive_bonus, reward_impact=-quad_impact_cost)
 

gym/envs/mujoco/humanoidstandup.py

@@ -2,18 +2,13 @@
 from gym.envs.mujoco import mujoco_env
 from gym import utils
 
-def mass_center(model):
-    mass = model.body_mass
-    xpos = model.data.xipos
-    return (np.sum(mass * xpos, 0) / np.sum(mass))[0]
-
 class HumanoidStandupEnv(mujoco_env.MujocoEnv, utils.EzPickle):
     def __init__(self):
         mujoco_env.MujocoEnv.__init__(self, 'humanoidstandup.xml', 5)
         utils.EzPickle.__init__(self)
 
     def _get_obs(self):
-        data = self.model.data
+        data = self.sim.data
         return np.concatenate([data.qpos.flat[2:],
                                data.qvel.flat,
                                data.cinert.flat,
@@ -23,8 +18,8 @@ def _get_obs(self):
 
     def _step(self, a):
         self.do_simulation(a, self.frame_skip)
-        pos_after = self.model.data.qpos[2][0]
-        data = self.model.data
+        pos_after = self.sim.data.qpos[2]
+        data = self.sim.data
         uph_cost = (pos_after - 0) / self.model.opt.timestep
 
         quad_ctrl_cost = 0.1 * np.square(data.ctrl).sum()

gym/envs/mujoco/inverted_double_pendulum.py

@@ -11,22 +11,22 @@ def __init__(self):
     def _step(self, action):
         self.do_simulation(action, self.frame_skip)
         ob = self._get_obs()
-        x, _, y = self.model.data.site_xpos[0]
+        x, _, y = self.sim.data.site_xpos[0]
         dist_penalty = 0.01 * x ** 2 + (y - 2) ** 2
-        v1, v2 = self.model.data.qvel[1:3]
+        v1, v2 = self.sim.data.qvel[1:3]
         vel_penalty = 1e-3 * v1**2 + 5e-3 * v2**2
         alive_bonus = 10
-        r = (alive_bonus - dist_penalty - vel_penalty)[0]
+        r = alive_bonus - dist_penalty - vel_penalty
         done = bool(y <= 1)
         return ob, r, done, {}
 
     def _get_obs(self):
         return np.concatenate([
-            self.model.data.qpos[:1],  # cart x pos
-            np.sin(self.model.data.qpos[1:]),  # link angles
-            np.cos(self.model.data.qpos[1:]),
-            np.clip(self.model.data.qvel, -10, 10),
-            np.clip(self.model.data.qfrc_constraint, -10, 10)
+            self.sim.data.qpos[:1],  # cart x pos
+            np.sin(self.sim.data.qpos[1:]),  # link angles
+            np.cos(self.sim.data.qpos[1:]),
+            np.clip(self.sim.data.qvel, -10, 10),
+            np.clip(self.sim.data.qfrc_constraint, -10, 10)
         ]).ravel()
 
     def reset_model(self):

gym/envs/mujoco/inverted_pendulum.py

@@ -22,7 +22,7 @@ def reset_model(self):
         return self._get_obs()
 
     def _get_obs(self):
-        return np.concatenate([self.model.data.qpos, self.model.data.qvel]).ravel()
+        return np.concatenate([self.sim.data.qpos, self.sim.data.qvel]).ravel()
 
     def viewer_setup(self):
         v = self.viewer