RoboPocket: Improve Robot Policies Instantly with Your Phone

Summary

RoboPocket is a robotics research paper (arXiv:2603.05504) by Junjie Fang, Wendi Chen, Han Xue, and colleagues, spanning robotics, AI, and machine learning. The work addresses a key limitation in scaling imitation learning: the efficiency of data collection. While handheld interfaces have become a scalable way to gather in-the-wild demonstrations, they typically operate open-loop — operators collect data blindly, unaware of the current policy's weaknesses, which results in inefficient coverage of critical state distributions. In contrast, interactive methods like DAgger can correct covariate shift but depend on physical setups that limit scalability. RoboPocket proposes using a phone as an interactive interface so operators can receive feedback about a policy's failure modes and instantly collect targeted corrective data, improving robot policies on the fly. The full abstract and paper are available at https://arxiv.org/abs/2603.05504.