Robot learning improves by adapting and handling fixed demo data effectively
How to Better Train VLAs: Lessons Learned From the REAL-I Challenge at ICRA 2026
Robotics
Summary
Teaching robots to do tasks using only a fixed set of example demonstrations is difficult. The authors looked at a competition where teams trained robots on a fixed budget of demonstrations and tested how well they performed tasks on real robots. The teams combined pretrained models with special ways to adapt to the real robot and keep previous skills, showing that how you handle the demonstration data and adjust to the deployment environment really matters. They also found that common ways to guess robot success from offline data don’t always predict real performance well.
What this means in practice
- •For industrial robotics teams: Design robot training processes that effectively adapt fixed demonstration data for better task success in real-world factory settings.
- •For robotics system integrators: Develop training pipelines combining pretrained vision-language-action models and adaptation methods for dual-arm humanoid robots to improve real task execution.
Authors
Jiaming Wang, Jizhuo Chen, Diwen Liu, Wang Song, Qiang Wang, Jie Ren, Chao Fu, Dingkun Zhu, Minchi Ruan, Hongtong Li, Yuhua Jiang, Zhiwei Xue, Yongping Pan, Harold Soh
Abstract
How can robot policies learn more effectively from a fixed demonstration budget? The first Real-world Embodied AI Learning (REAL-I) Challenge at ICRA 2026 examined this question through simulation, real-robot evaluation, and an on-site final on a shared dual-arm humanoid platform. We describe the challenge tasks, data and deployment interfaces, and competition results, then compare the approaches contributed by NUS-CLEAR, RCL-Lab, and Deeptouch.ai. Their systems combined pretrained vision-language-action models and task-specific imitation policies with different strategies for data curation, staged adaptation, checkpoint selection, and action-space design. The team reports highlight the importance of adapting to the deployment environment while retaining prior capabilities, treating demonstration quality at an appropriate temporal scale, and suppressing errors in inactive robot components. They also expose the limitations of offline action-prediction metrics for forecasting closed-loop success. These observations motivate a view of fixed-data robot learning that integrates data, adaptation, evaluation, and deployment.