Difference-Aware Retrieval Policies for Imitation Learning
arXiv·low signal
Parametric imitation learning via behavior cloning generalizes poorly to out-of-distribution states because errors compound across a trajectory. This work proposes difference-aware retrieval policies that surface relevant demonstrations at inference rather than relying solely on a fixed parametric model. It is relevant to builders of robotic and embodied agents seeking more robust, recoverable imitation-learned behavior.