---
license: mit
tags:
  - robotics
  - imitation-learning
  - rlds
  - openvla
  - libero
  - boss-benchmark
---

# BOSS — RAMG-augmented demonstrations (RLDS)

The augmented training set behind Setup B of Table I in
[**BOSS**](https://arxiv.org/abs/2502.15679) (Yang et al., *IEEE RA-L*, 2025), in
RLDS/TFDS form so it can be fed
straight to OpenVLA fine-tuning.

- **56,945 episodes** across 1,727 modified tasks (the 44 BOSS skills, each with
  one visual modification generated by the Rule-based Automatic Modification
  Generator)
- 1,024 tfrecord shards, ~267 GB

Produced by replaying each original demonstration in its modified environment and
discarding failed replays, so the trajectory is unchanged and only the visual
observation differs.

## Usage

```shell
huggingface-cli download yygx/BOSS-data-augmentation --repo-type dataset \
  --local-dir datasets

# from inside your OpenVLA checkout
DATASET_NAME=libero_bl3_all DATA_ROOT_DIR=./datasets \
  bash ../integrations/openvla/shells/finetune_openvla.sh
```

## HDF5 form

The BC baselines read HDF5, not RLDS, and that form is **not** distributed here.
Regenerate it from the repository, which ships all 1,727 modified bddl files:

```shell
python RAMG/DA_demos_generation.py --benchmark data_augmentation
```

Roughly 6 minutes per task single-threaded; `--start-index N` resumes.

## Citation

```bibtex
@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}
}
```