Skip to content

Conversation

s1lent4gnt
Copy link
Contributor

@s1lent4gnt s1lent4gnt commented Sep 1, 2025

What this does

Implement ConRFT (Consistency-based Reinforced Fine-Tuning) approach for fine-tuning Vision-Language-Action (VLA) models in robotic manipulation tasks.

NOTE: this PR depends on #1831

How to test it

Cal-ConRFT (offline)

python src/lerobot/scripts/rl/learner.py --config json/train_conrft_offline.json

HIL-ConRFT (online)

python src/lerobot/scripts/rl/learner.py --config json/train_conrft_online_learner.json
python src/lerobot/scripts/rl/actor.py --config json/train_conrft_online_actor.json

You can find config files in this branch lilkm/configs here : https://github.com/s1lent4gnt/lerobot/tree/lilkm/configs/json

TODO

  • Implement state stacking and masking observation in OctoEncodingWrapper.
  • Add/Implement mc_returns in dataset.
  • Investigate slow training in offline phase.

This module implements the ConRFT (Consistency-based Reinforced Fine-Tuning) approach
for fine-tuning Vision-Language-Action (VLA) models in robotic manipulation tasks.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

1 participant