ConRFT: A Reinforced Fine-tuning Method for VLA Models via Consistency Policy
Fuente:
arXiv
Gespeichert in:
| Hauptverfasser: | , , , , , |
|---|---|
| Format: | Preprint |
| Veröffentlicht: |
2025
|
| Schlagworte: | |
| Online-Zugang: | |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| _version_ | 1866910910277222400 |
|---|---|
| author | Chen, Yuhui Tian, Shuai Liu, Shugao Zhou, Yingting Li, Haoran Zhao, Dongbin |
| author_facet | Chen, Yuhui Tian, Shuai Liu, Shugao Zhou, Yingting Li, Haoran Zhao, Dongbin |
| contents | Vision-Language-Action (VLA) models have shown substantial potential in real-world robotic manipulation. However, fine-tuning these models through supervised learning struggles to achieve robust performance due to limited, inconsistent demonstrations, especially in contact-rich environments. In this paper, we propose a reinforced fine-tuning approach for VLA models, named ConRFT, which consists of offline and online fine-tuning with a unified consistency-based training objective, to address these challenges. In the offline stage, our method integrates behavior cloning and Q-learning to effectively extract policy from a small set of demonstrations and stabilize value estimating. In the online stage, the VLA model is further fine-tuned via consistency policy, with human interventions to ensure safe exploration and high sample efficiency. We evaluate our approach on eight diverse real-world manipulation tasks. It achieves an average success rate of 96.3% within 45-90 minutes of online fine-tuning, outperforming prior supervised methods with a 144% improvement in success rate and 1.9x shorter episode length. This work highlights the potential of integrating reinforcement learning to enhance the performance of VLA models for real-world robotic applications. Videos and code are available at our project website https://cccedric.github.io/conrft/. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2502_05450 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | ConRFT: A Reinforced Fine-tuning Method for VLA Models via Consistency Policy Chen, Yuhui Tian, Shuai Liu, Shugao Zhou, Yingting Li, Haoran Zhao, Dongbin Robotics Artificial Intelligence Vision-Language-Action (VLA) models have shown substantial potential in real-world robotic manipulation. However, fine-tuning these models through supervised learning struggles to achieve robust performance due to limited, inconsistent demonstrations, especially in contact-rich environments. In this paper, we propose a reinforced fine-tuning approach for VLA models, named ConRFT, which consists of offline and online fine-tuning with a unified consistency-based training objective, to address these challenges. In the offline stage, our method integrates behavior cloning and Q-learning to effectively extract policy from a small set of demonstrations and stabilize value estimating. In the online stage, the VLA model is further fine-tuned via consistency policy, with human interventions to ensure safe exploration and high sample efficiency. We evaluate our approach on eight diverse real-world manipulation tasks. It achieves an average success rate of 96.3% within 45-90 minutes of online fine-tuning, outperforming prior supervised methods with a 144% improvement in success rate and 1.9x shorter episode length. This work highlights the potential of integrating reinforcement learning to enhance the performance of VLA models for real-world robotic applications. Videos and code are available at our project website https://cccedric.github.io/conrft/. |
| title | ConRFT: A Reinforced Fine-tuning Method for VLA Models via Consistency Policy |
| topic | Robotics Artificial Intelligence |
| url | https://arxiv.org/abs/2502.05450 |