Towards Online Adaptation for Autonomous Household Assistants

Many assistive home robotics applications assume open-loop interactions: robots incorporate little feedback from people while autonomously completing tasks. This places undue burden on people to condition their actions and environment to maximize the likelihood of their desired outcomes. We formalize assistive household rearrangement as collaborative online inverse reinforcement learning (IRL). Since online IRL can lead to sample inefficient interactions and overfit to specific user objectives, we compare sample efficiency and generalizability of two initial choices of action representations in a simulated household rearrangement task. We show, under certain assumptions, that representing objects by their material properties can increase sample efficiency and generalizability to out of domain objects.