libero-rnbencore-models

LIBERO R&B-EnCoRe Models

Part of the R&B-EnCoRe work from the Stanford Autonomous Systems Lab.

• 📄 Paper: https://arxiv.org/abs/2602.08167
• 💻 Code / usage: https://github.com/rnb-encore/RnB-EnCoRe-Manipulation

Three vision-language-action (VLA) policies fine-tuned on the LIBERO manipulation benchmark, built on a Qwen2.5 backbone. The checkpoints differ in how — and whether — they perform embodied reasoning before predicting actions. All three are packaged in this single repository, each under its own subfolder.

Models

Each subfolder is self-contained:

<subfolder>/
├── checkpoints/
│   └── latest-checkpoint.pt
├── config.json
└── dataset_statistics.json

Model details

• Base model: Qwen2.5-based VLA
• Modality: image + text → robot action tokens
• Task: language-conditioned manipulation on LIBERO
• Output: optional embodied-reasoning trace followed by low-level actions
• Checkpoint format: PyTorch .pt at <subfolder>/checkpoints/latest-checkpoint.pt, with companion config.json and dataset_statistics.json

Usage

These are .pt checkpoints loaded through the project codebase rather than the transformers API — download the model folder, then point the codebase at the checkpoint.

Install the CLI:

pip install -U huggingface_hub

Download the whole repository (all three models):

hf download stanfordasl/libero-rnbencore-models --local-dir ./libero-rnbencore-models

Or download a single model by filtering to its subfolder:

hf download stanfordasl/libero-rnbencore-models \
  --include "libero-qwen25-openvla/*" \
  --local-dir ./libero-rnbencore-models

Point your codebase at the checkpoint:

# Replace the entrypoint/flag with the project codebase's eval script.
python run_libero_eval.py \
  --pretrained_checkpoint ./libero-rnbencore-models/libero-qwen25-openvla/checkpoints/latest-checkpoint.pt

Swap the subfolder name for libero-qwen25-object-full-reasoning-vla or libero-qwen25-object-rnbencore-reasoning-vla to run the other policies. For dataset prep, prompting, inference, and evaluation, follow the project repository above.

Intended use & limitations

These checkpoints are research artifacts trained on the LIBERO simulation benchmark. They are not intended for deployment on physical hardware without further validation; behavior outside the LIBERO task distribution is untested. Always validate in simulation before any downstream use.

Citation

If you use these models, please cite the RnB-EnCoRe work:

• Paper: https://arxiv.org/abs/2602.08167
• Code: https://github.com/rnb-encore/RnB-EnCoRe-Manipulation