---
datasets:
- behavior-1k/2025-challenge-demos
pipeline_tag: robotics
license: apache-2.0
tags:
- embodied-ai
- vla
- behavior-1k
---

# Openpi Comet: Competition Solution for 2025 BEHAVIOR Challenge

This repository contains the model weights of Team Comet for the [2025 BEHAVIOR Challenge](https://behavior.stanford.edu/index.html). Our [submission](https://behavior.stanford.edu/challenge/leaderboard.html#privileged-information-track) achieved a Q-score of 0.2514, securing 2nd place overall and finishing just behind the winning team by a narrow margin—highlighting both the strong competitiveness of our approach and the effectiveness of our end-to-end Vision-Language-Action (VLA) training strategy.

- **Paper:** [Openpi Comet: Competition Solution For 2025 BEHAVIOR Challenge](https://arxiv.org/abs/2512.10071)
- **Repository:** [https://github.com/mli0603/openpi-comet](https://github.com/mli0603/openpi-comet)
- **Project Page:** [BEHAVIOR Challenge](https://behavior.stanford.edu/index.html)

## Model Description

OpenPi Comet is a foundation model designed for long-horizon mobile manipulation tasks in realistic household settings, specifically developed for the BEHAVIOR-1K dataset. Building on $\pi_{0.5}$ (Pi05), the model incorporates hierarchical instructions (global, subtask, skill) and multimodal observations (RGB, depth, point cloud, etc.). Through systematic training techniques and data scaling, including Rejection Sampling Fine-Tuning (RFT), the team achieved a validation Q-score of 0.345.

## Usage

Please refer to [the official GitHub repository](https://github.com/mli0603/openpi-comet) for detailed installation and usage instructions. The codebase provides a unified framework for:
- Distributed pre-training and fine-tuning of OpenPi models.
- Data generation (teleoperation and simulation rollouts).
- Post-training via Rejection Sampling Fine-Tuning (RFT).
- Evaluation within the BEHAVIOR-1K simulator.

### Quick Setup
```bash
git clone https://github.com/mli0603/openpi-comet.git
cd openpi-comet
uv sync
uv pip install -e .
```

## Citation

If you find this work useful, please consider citing:

```bibtex
@article{bai2025openpicometcompetitionsolution,
  title={Openpi Comet: Competition Solution For 2025 BEHAVIOR Challenge}, 
  author={Junjie Bai and Yu-Wei Chao and Qizhi Chen and Jinwei Gu and Moo Jin Kim and Zhaoshuo Li and Xuan Li and Tsung-Yi Lin and Ming-Yu Liu and Nic Ma and Kaichun Mo and Delin Qu and Shangkun Sun and Hongchi Xia and Fangyin Wei and Xiaohui Zeng},
  journal={arXiv preprint arXiv:2512.10071},
  year={2025},
  url={https://arxiv.org/abs/2512.10071}, 
}
```