Demonstration of the EMPATHIC Framework for Task Learning from Implicit Human Feedback

Reactions such as gestures, facial expressions, and vocalizations are an abundant, naturally occurring channel of information that humans provide during interactions. An agent could leverage understanding implicit human feedback to improve its task performance at no cost the human. This approach contrasts with common teaching methods based on demonstrations, critiques, or other guidance need be attentively intentionally provided. In this work, we demonstrate a novel data-driven framework for learning from feedback, EMPATHIC. two-stage method consists (1) mapping relevant statistics reward, optimality, advantage; (2) using learn task. We instantiate first stage three second-stage evaluations learned mapping. To do so, collect dataset reactions while participants observe execute sub-optimal policy prescribed training train deep neural network data ability infer relative reward ranking events in prerecorded reactions; live (3) transfer domain which it evaluates robot manipulation trajectories. video, focus demonstrating online capability our instantiation