rlworkgroup / garage

A toolkit for reproducible reinforcement learning research.
MIT License
1.88k stars 310 forks source link

Kernels restarting on specific environments #1508

Closed cversteeg closed 4 years ago

cversteeg commented 4 years ago

Thanks for updating this package; very useful! I am encountering an issue where many of the environments cause my kernel to crash. I am able to run mountain-car and cartpole without issues, but the inverted pendulums and any mujoco models cause my kernel to crash without any indication of the cause.

Specifically, this line causes it to crash env = TfEnv(gym.make('InvertedDoublePendulum-v2')) while this does not. env = TfEnv(env_name='MountainCarContinuous-v0')

Any advice would be appreciated. Thanks!

cversteeg commented 4 years ago

Ok, update; running the script inside of the terminal gives some more information.

ERROR: Invalid text in activation key

I assume that the mujoco activation key is incorrect, which is causing it to fail silently. I followed the installation instructions and included the path to my key in the installation step. Any way to check to see why this is failing?

Thanks!

ryanjulian commented 4 years ago

@cversteeg I should note that Jupyter/colab are not officially-supported environments (i.e. we don't test them in the CI) so any support for those will be best-effort.

in this case it's likely that you're missing a license key for MuJoCo, which is required to run those environments.

ryanjulian commented 4 years ago

You can start by verifying that your mujoco key is present and uncorrupted at ~/.mujoco/mjkey.txt after the installation but before you try to run an experiment

cversteeg commented 4 years ago

Yep!

The issue was that I made a placeholder mujoco key to just get the basics working while I waited for the license file from Mujoco. The setup apparently didn't overwrite that file when I went back to reinstall the mujoco portion.

Thanks so much for the prompt response!

ryanjulian commented 4 years ago

You're welcome!