Reproduce “Beyond Imitation: Reinforcement Learning-Based Sim-Real Co-Training for VLA Models” 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

Liangzhi Shi, Shuaihang Chen, Feng Gao, Yinuo Chen, Kang Chen, Tonghe Zhang, Hongzhi Zang, Jiakai Zhou and 3 more

RL-Co trains vision-language-action policies first on real and simulated demonstrations, then through reward-guided interaction in simulation while retaining supervision from real demonstrations. This combines supervised fine-tuning with reinforcement learning. In four-task real-robot evaluation, OpenVLA achieves a macro-average success rate of 64.0%, versus 40.0% for supervised co-training. Results cover one robot embodiment and two evaluation sessions.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.