ypz21/HARP_VLA_calvin

HARP-VLA — CALVIN

Merged HARP-VLA policy for CALVIN evaluation.

Project website · Code

Use the official inference code to load this checkpoint. The vision adapters are included in the merged weights and installed by the HARP-VLA model loader.

python examples/download_checkpoint.py
python -m experiments.robot.calvin.evaluate \
  --pretrained_checkpoint checkpoints/calvin \
  --calvin_root "$CALVIN_DATA_ROOT"

The policy consumes static and gripper camera images and seven-dimensional proprioception. It predicts ten seven-dimensional relative actions per inference call. Use center cropping and the supplied normalization statistics.

Acknowledgements

Built upon OpenVLA-OFT and UniVLA. Evaluation uses CALVIN.

Citation

@article{zhu2026harp,
  title={HARP-VLA: Human-Robot Aligned Representation Learning for Vision-Language-Action Model},
  author={Zhu, Xiang and Yuan, Puzhen and Liu, Yichen and Chen, Jianyu},
  journal={arXiv preprint arXiv:2605.31234},
  year={2026}
}