Reproduce “How to Better Train VLAs: Lessons Learned From the REAL-I Challenge at ICRA 2026” on OpenResearch

About this paper

Jiaming Wang, Jizhuo Chen, Diwen Liu, Wang Song, Qiang Wang, Jie Ren, Chao Fu, Dingkun Zhu and 6 more

The REAL-I Challenge at ICRA 2026 compares fixed-data strategies for deploying vision-language-action policies on dual-arm humanoid robots, showing that performance depends strongly on adaptation, demonstration curation, action design, and closed-loop validation—not only on model scale. NUS-CLEAR achieved the highest overall score of 71.0, followed by RCL-Lab at 69.1, while task-specific policies were especially competitive on dexterous manipulation.

Key insights

• Small amounts of deployment-specific adaptation can substantially improve transfer, but unrestricted fine-tuning on limited data risks overfitting and forgetting earlier capabilities.
• Demonstration quality and task structure matter: duration-based filtering can remove incomplete or inefficient episodes, while rare-event selection and oversampling can improve behaviors such as parcel reorientation.
• A single large VLA is not uniformly optimal. Unified models performed strongly on structured tasks, whereas a specialized Diffusion Policy achieved a perfect score on the bimanual bottle task; masking inactive arms also reduced drift and interference.

Summary by alphaXiv.