Imitation learning policies often struggle when deployed in closed-loop settings, where tiny observation or action errors accumulate and push the system off course. A new arXiv paper, "Diffeomorphic Hidden-State Steering," investigates a counterintuitive fix: using the original expert demonstrations at test time to keep the policy on track.
The authors propose steering the policy's hidden states during execution, guided by the geometric structure of expert demonstrations. Their approach, called diffeomorphic hidden-state steering, aims to make the policy more robust to the small perturbations that normally compound into failures.
The paper is framed as a study of whether expert data can be repurposed beyond training, as a test-time corrective signal. While the abstract does not report specific results, it suggests a new direction for improving imitation learning robustness without altering the training procedure itself.