Reproduce “ExToken: Structured Exploration for Efficient Vision-Language-Action Reinforcement Fine-tuning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yilun Kong, Yunpeng Qing, Guozheng Ma, Haoyu Wang, Li Shen, Zhi Hou, Dacheng Tao
ExToken is a framework designed to enhance the sample efficiency of Vision-Language-Action Reinforcement Learning (VLA-RL) by promoting structured exploration. It uses exploration tokens derived from offline demonstrations to guide the policy, resulting in improved task performance, faster convergence, and better robustness in both simulated and real-world robotic manipulation tasks.
Key insights
- Trajectory diversity is more critical for VLA-RL sample efficiency and algorithmic performance than the sheer quantity of raw interaction data collected.
- Standard stochastic exploration often causes 'action mode collapse' in VLA-RL, leading policies to repeatedly exploit similar action patterns and plateau at suboptimal performance.
- Discrete behavioral priors extracted from offline demonstrations can effectively provide structured guidance for exploration, alleviating the observed exploration stagnation in VLA-RL.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.