Skip to content

Conversation

s1lent4gnt
Copy link
Contributor

@s1lent4gnt s1lent4gnt commented Sep 1, 2025

What this does

Implement Octo VLA policy.

How to fine-tune with LeRobot

lerobot-train --policy.type=octo --dataset.repo_id=lilkm/panda_pick_octo_resized --policy.train_action_head_only=True

TODO

  • Make Octo observation naming LeRobot native.
  • Add other fine-tuning configurations, like adding proprio.

@s1lent4gnt s1lent4gnt mentioned this pull request Sep 1, 2025
3 tasks
@lukicdarkoo
Copy link

lukicdarkoo commented Sep 2, 2025

Nice work! Did you measure performance? How does it compare to other LeRobot policies?

@s1lent4gnt
Copy link
Contributor Author

Thanks @lukicdarkoo, no I didn't measure it performance against other VLAs in LeRobot. I will do it at some point.

- Replaced nested for-loop attention mask  creation with fully vectorized PyTorch implementation.
- Removed the `TokenMetadata` class from `base.py`.
- Introduced a `RULE_MAP` to convert attention rules to integers for faster processing.
@jadechoghari jadechoghari added enhancement Suggestions for new features or improvements policies Items related to robot policies labels Sep 23, 2025
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
enhancement Suggestions for new features or improvements policies Items related to robot policies
Projects
None yet
Development

Successfully merging this pull request may close these issues.

3 participants