lerobot
lerobot copied to clipboard
Fix gymnasium version as pre-1.0.0
What this does
This PR fixes #470 (🐛 Bug)
The issue is caused by recent v1.0.0 release of gymnasium
from 08.10.2024.
It looks like the final_info
key-value pair is no longer available in the output of SyncVectorEnv.step
, so evaluation code for success_rate in lerobot/scripts/eval.py gets broken.
The proper fix of the issue requires changes in eval.py
and subsequent checks versus available configs.
This PR fixes evaluation behavior "as it was before" and postpones the usage of gymnasium==1.0.0 until necessary changes and checks are done.
How it was tested
I have installed gymnasium==0.29.1
(which is the latest pre-1.0.0) locally and re-evaluated trained diffusion policy for pushT and achieved correct pc_success
:
{'avg_sum_reward': 101.66934858131904, 'avg_max_reward': 0.9727046033094218, 'pc_success': 76.0, 'eval_s': 69.3367211818695, 'eval_ep_s': 1.3867344379425048}
How to checkout & try? (for the reviewer)
git remote add zimka https://github.com/zimka/lerobot
git checkout zimka fix_gymnasium_version