Improving Vision-Language-Action Model with Online Reinforcement Learning
This work addresses the challenge of fine-tuning large VLA models for robotics, offering a method to enhance performance during environment interaction, though it is incremental as it builds on existing VLA and RL techniques.
The paper tackles the problem of improving vision-language-action (VLA) models for robotic control by addressing training instability and computational burdens when applying online reinforcement learning, proposing an iterative framework that combines reinforcement and supervised learning, which is validated in simulated and real-world benchmarks.
Recent studies have successfully integrated large vision-language models (VLMs) into low-level robotic control by supervised fine-tuning (SFT) with expert robotic datasets, resulting in what we term vision-language-action (VLA) models. Although the VLA models are powerful, how to improve these large models during interaction with environments remains an open question. In this paper, we explore how to further improve these VLA models via Reinforcement Learning (RL), a commonly used fine-tuning technique for large models. However, we find that directly applying online RL to large VLA models presents significant challenges, including training instability that severely impacts the performance of large models, and computing burdens that exceed the capabilities of most local machines. To address these challenges, we propose iRe-VLA framework, which iterates between Reinforcement Learning and Supervised Learning to effectively improve VLA models, leveraging the exploratory benefits of RL while maintaining the stability of supervised learning. Experiments in two simulated benchmarks and a real-world manipulation suite validate the effectiveness of our method.