Kill timestep_limit

examples/scripts/sim_env

@@ -20,7 +20,7 @@ env = envs.make(args.env)
 ac_space = env.action_space
 
 fps = args.fps or env.metadata.get('video.frames_per_second') or 100
-if args.max_steps == 0: args.max_steps = env.spec.timestep_limit
+if args.max_steps == 0: args.max_steps = env.spec.tags['wrapper_config.TimeLimit.max_episode_steps']
 
 while True:
     env.reset()

gym/envs/README.md

@@ -56,12 +56,10 @@ ant.AntEnv
   register(
       id='foo-v0',
       entry_point='gym_foo.envs:FooEnv',
-      timestep_limit=1000,
   )
   register(
       id='foo-extrahard-v0',
       entry_point='gym_foo.envs:FooExtraHardEnv',
-      timestep_limit=1000,
   )
   ```
 

gym/envs/registration.py

@@ -25,7 +25,6 @@ class EnvSpec(object):
     Args:
         id (str): The official environment ID
         entry_point (Optional[str]): The Python entrypoint of the environment class (e.g. module.name:Class)
-        timestep_limit (int): The max number of timesteps per episode during training
         trials (int): The number of trials to average reward over
         reward_threshold (Optional[int]): The reward threshold before the task is considered solved
         local_only: True iff the environment is to be used only on the local machine (e.g. debugging envs)
@@ -35,14 +34,12 @@ class EnvSpec(object):
 
     Attributes:
         id (str): The official environment ID
-        timestep_limit (int): The max number of timesteps per episode in official evaluation
         trials (int): The number of trials run in official evaluation
     """
 
-    def __init__(self, id, entry_point=None, timestep_limit=1000, trials=100, reward_threshold=None, local_only=False, kwargs=None, nondeterministic=False, tags=None):
+    def __init__(self, id, entry_point=None, trials=100, reward_threshold=None, local_only=False, kwargs=None, nondeterministic=False, tags=None):
         self.id = id
         # Evaluation parameters
-        self.timestep_limit = timestep_limit
         self.trials = trials
         self.reward_threshold = reward_threshold
         # Environment properties

misc/write_rollout_data.py

@@ -13,13 +13,13 @@ def __init__(self, ac_space):
     def act(self, _):
         return self.ac_space.sample()
 
-def rollout(env, agent, timestep_limit):
+def rollout(env, agent, max_episode_steps):
     """
-    Simulate the env and agent for timestep_limit steps
+    Simulate the env and agent for max_episode_steps
     """
     ob = env.reset()
     data = collections.defaultdict(list)
-    for _ in xrange(timestep_limit):
+    for _ in xrange(max_episode_steps):
         data["observation"].append(ob)
         action = agent.act(ob)
         data["action"].append(action)
@@ -46,10 +46,10 @@ def main():
     alldata = {}
     for i in xrange(2):
         np.random.seed(i)
-        data = rollout(env, agent, env.spec.timestep_limit)
+        data = rollout(env, agent, env.spec.tags['wrapper_config.TimeLimit.max_episode_steps'])
         for (k, v) in data.items():
             alldata["%i-%s"%(i, k)] = v
     np.savez(args.outfile, **alldata)
 
 if __name__ == "__main__":
-    main()
+    main()