admin/lerobot
1
0
Fork 0
mirror of https://github.com/huggingface/lerobot.git synced 2026-06-18 16:57:12 +00:00
Code Issues Packages Projects Releases Wiki Activity

Add FastWAM policy review updates

Browse Source
This commit is contained in:
ZibinDong
2026-06-09 05:34:13 +00:00
committed by Maxime Ellerbach
parent a343ed3a63
commit dfc0170b4d
23 changed files with 671 additions and 3597 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/source/fastwam.mdx
-16
View File
@@ -128,9 +128,6 @@ evaluation pipeline:
--policy.toggle_action_dimensions='[-1]'
```
`policy.invert_dimensions` remains available for older checkpoints or robot
setups that only need a sign inversion.
## Results
Evaluated on LIBERO with [`ZibinDong/fastwam_libero_uncond_2cam224`](https://huggingface.co/ZibinDong/fastwam_libero_uncond_2cam224):
@@ -145,19 +142,6 @@ Evaluated on LIBERO with [`ZibinDong/fastwam_libero_uncond_2cam224`](https://hug
Reproduce: `lerobot-eval --policy.path=ZibinDong/fastwam_libero_uncond_2cam224 --policy.device=cuda --policy.torch_dtype=float32 --policy.n_action_steps=10 --env.type=libero --env.task=libero_spatial --env.observation_height=256 --env.observation_width=256 --eval.batch_size=1 --eval.n_episodes=50 --seed=0 --env.episode_length=300` (1x H20 140 GB).
## Reproducibility Checklist
For a PR adding or updating FastWAM results, include:
- the training dataset repo id
- the LeRobot-format checkpoint repo id
- the exact `lerobot-train` command
- the exact `lerobot-eval` or `lerobot-rollout` command
- the number of evaluation episodes
- the GPU type and count
The upstream Fast-WAM release provides reference checkpoints and benchmark assets at `yuanty/fastwam`; LeRobot eval numbers should be reported from a converted LeRobot-format checkpoint so reviewers can reproduce them with the commands above.
## References
- [Fast-WAM paper](https://arxiv.org/abs/2603.16666)