yygx/BOSS-openvla-adapters

BOSS — fine-tuned OpenVLA models

OpenVLA-7B fine-tuned for the BOSS benchmark, covering both setups of Table I in the paper (IEEE RA-L, 2025).

Folder	Paper	Fine-tuned on
setup_a_libero44/	Table I, Setup A	the 44 original BOSS skills
setup_b_augmented/	Table I, Setup B	the RAMG-augmented set (56,945 episodes)

LoRA rank 32, batch size 8, lr 5e-4, image augmentation on. The weights are merged, so each folder is a full model, not an adapter.

Usage

huggingface-cli download yygx/BOSS-openvla-adapters --include "setup_a_libero44/*" \
  --local-dir runs

# from inside your OpenVLA checkout, after
# `python <boss>/integrations/openvla/install.py`
SEED=10000 bash ../integrations/openvla/shells/eval_openvla.sh

Citation

@article{yang2025boss,
  title={BOSS: Benchmark for observation space shift in long-horizon task},
  author={Yang, Yue and Zhao, Linfeng and Ding, Mingyu and Bertasius, Gedas and Szafir, Daniel},
  journal={IEEE Robotics and Automation Letters},
  volume={10},
  number={9},
  pages={8882--8889},
  year={2025},
  publisher={IEEE}
}

Built on OpenVLA and LIBERO.