Ergodic imitation method lets robots adaptively explore beyond demonstrations

Imitation learning often fails when robots encounter mismatches between training and deployment environments. Ziyi Xu, Cem Bilaloglu, Yiming Li, and Sylvain Calinon introduce a novel ergodic imitation method that allows robots to adaptively explore around demonstrations. Their approach builds a target distribution from the geometry of retrieved demonstrations and uses it within a receding-horizon control framework. This enables the robot to smoothly interpolate between tracking the nominal trajectory and exploring new regions, preventing it from getting stuck when environmental changes or imperfect observations arise.

The paper extends ergodic control—traditionally used for area coverage and search—into the realm of imitation learning. By grounding exploration in the demonstrated behavior, the method maintains task relevance while adapting online. Published as a 4-page arXiv paper with 3 figures, this work offers a practical solution for deploying robots in dynamic, real-world settings. Potential applications include manufacturing, service robotics, and any domain where robots must handle variability without full retraining.