search

google-deepmind / dm_control

Google DeepMind's software stack for physics-based simulation and Reinforcement Learning environments, using MuJoCo.
Apache License 2.0
3.72k stars 660 forks source link

dm_control.rl.control.PhysicsError: Physics state is invalid. Warning(s) raised: mjWARN_CONTACTFULL #220

Open 13253591602 opened 2 years ago

13253591602 commented 2 years ago

Hi, I am trying to reproduce the experiment from paper "Learning human behaviors from motion capture by adversarial imitation". I used the code from https://github.com/huiwenzhang/merel-mocap-gail. But after running about 400 iterations, it throws the following error: WARNING:absl:Pre-allocated contact buffer is full. Increase nconmax above 100. Time = 0.0000. Traceback (most recent call last): File "/home/yh/anaconda3/envs/gail/lib/python3.7/runpy.py", line 193, in _run_module_as_main "main", mod_spec) File "/home/yh/anaconda3/envs/gail/lib/python3.7/runpy.py", line 85, in _run_code exec(code, run_globals) File "/home/yh/Projects/IL/merel-mocap-gail-master/baselines/baselines/gail/run_mujoco.py", line 268, in main(args) File "/home/yh/Projects/IL/merel-mocap-gail-master/baselines/baselines/gail/run_mujoco.py", line 125, in main args.obs_only File "/home/yh/Projects/IL/merel-mocap-gail-master/baselines/baselines/gail/run_mujoco.py", line 181, in train obs_only=obs_only) File "/home/yh/Projects/IL/merel-mocap-gail-master/baselines/baselines/gail/trpo_mpi.py", line 316, in learn seg = seg_gen.next() File "/home/yh/Projects/IL/merel-mocap-gail-master/baselines/baselines/gail/trpo_mpi.py", line 132, in traj_segment_generator env.reset() File "/home/yh/anaconda3/envs/gail/lib/python3.7/site-packages/dm_control/rl/control.py", line 82, in reset self._task.initialize_episode(self._physics) File "/home/yh/anaconda3/envs/gail/lib/python3.7/site-packages/dm_control/suite/humanoid_CMU.py", line 136, in initialize_episode physics.after_reset() File "/home/yh/anaconda3/envs/gail/lib/python3.7/site-packages/dm_control/mujoco/engine.py", line 298, in after_reset self.forward() File "/home/yh/anaconda3/envs/gail/lib/python3.7/site-packages/dm_control/mujoco/engine.py", line 308, in forward mjlib.mj_forward(self.model.ptr, self.data.ptr) File "/home/yh/anaconda3/envs/gail/lib/python3.7/contextlib.py", line 119, in exit next(self.gen) File "/home/yh/anaconda3/envs/gail/lib/python3.7/site-packages/dm_control/mujoco/engine.py", line 331, in check_invalid_state raise _control.PhysicsError(message) dm_control.rl.control.PhysicsError: Physics state is invalid. Warning(s) raised: mjWARN_CONTACTFULL

I have tried all dm_control version.

Markus28 commented 2 years ago

Somewhat connected to https://github.com/deepmind/deepmind-research/issues/307

vaxenburg commented 2 years ago

This happens when the number of detected contacts is larger than the allocated contact buffer (as the error message says). It is explained, e.g., here. One way to increase the buffer size is to tweak the nconmax parameter in your model's XML file. You want to look for an entry that might look like <size nconmax="100"/>.