---
license: mit
tags:
  - robotics
  - imitation-learning
  - behavioral-cloning
  - libero
  - boss-benchmark
---

# BOSS — pre-trained skill policies

Skill policies for the three Behavioral Cloning baselines of
[**BOSS: Benchmark for Observation Space Shift in Long-Horizon Task**](https://arxiv.org/abs/2502.15679)
(Yang et al., *IEEE Robotics and Automation Letters*, 2025),
so the benchmark can be run without re-training.

Code: https://github.com/YY-GX/BOSS · Project page: https://boss-benchmark.github.io/

## Contents

| Folder | Baseline | Files |
|---|---|---|
| `BCRNNPolicy_seed10000/run_001/` | BC-RESNET-RNN | `task0..task43_model.pth` |
| `BCTransformerPolicy_seed10000/run_001/` | BC-RESNET-T | `task0..task43_model.pth` |
| `BCViLTPolicy_seed10000/run_001/` | BC-VIT-T | `task0..task43_model.pth` |

One checkpoint per skill, indexed by position in the `boss_44` task list. All
trained with `seed=10000`, observation space = third-person camera + 7-DoF joint
angles + 2-DoF gripper state (no wrist camera), matching Section V-A of the paper.

## Usage

```shell
huggingface-cli download yygx/BOSS-checkpoints --local-dir experiments/boss_44/0.0.0

python libero/lifelong/eval_skills_unaffected_by_oss.py \
  --benchmark boss_44 --seed 10000 --max_steps 400 \
  --model_path_folder experiments/boss_44/0.0.0/BCTransformerPolicy_seed10000/run_001/
```

## Important: `--max_steps 400`

Evaluation hyper-parameters are read from the checkpoint, and these checkpoints
carry `max_steps=600` while the published results used **400**. On BOSS-C3 task 1,
BC-RESNET-T reaches a Delta to Upper Bound Ratio of 89% at 400 steps — the value
in Figure 5 — and 67% at 600. Always pass `--max_steps 400` to reproduce the paper.

## 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}
}
```

Built on [LIBERO](https://github.com/Lifelong-Robot-Learning/LIBERO) (Liu et al., 2023).