Skip to content

feat(openpi): add aligned LIBERO training support - #1500

Open
Chernobyllight wants to merge 6 commits into
ModelTC:mainfrom
Chernobyllight:lightx2v-hunyuan-v17.5
Open

feat(openpi): add aligned LIBERO training support#1500
Chernobyllight wants to merge 6 commits into
ModelTC:mainfrom
Chernobyllight:lightx2v-hunyuan-v17.5

Conversation

@Chernobyllight

Copy link
Copy Markdown
Contributor

Add official-aligned OpenPI π0.5 LIBERO PyTorch fine-tuning with FP32 master weights, BF16 compute, EMA, strict checkpoint resume, and streamlined inference, ROS evaluation, and runtime setup.

Integrate OpenPI pi0.5-LIBERO with local PyTorch inference, checkpoint conversion, runtime setup, and official single- and multi-GPU evaluation workflows.
Integrate pi0.5-LIBERO inference with local PyTorch execution, official-compatible evaluation, multi-GPU suite scheduling, and decoupled ROS2 policy/simulator nodes. Simplify the integration while preserving preprocessing, FP64 actions, and continuous policy RNG semantics.
Add official-aligned PyTorch fine-tuning with FP32 master weights, BF16 compute, EMA, strict checkpoint resume, and LIBERO preprocessing.

Streamline OpenPI inference, ROS evaluation, and runtime validation while keeping generated artifacts out of Git.
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