Learning under Misspecified Objective Spaces

摘要: Learning robot objective functions from human input has become increasingly important, but state-of-the-art techniques assume that the human's desired lies within robot's hypothesis space. When this is not true, even methods keep track of uncertainty over fail because they reason about which might be correct, and whether any hypotheses are correct. We focus specifically on learning physical corrections during task execution, where having a rich enough space leads to updating its in ways person did actually intend. observe such appear irrelevant robot, best way achieving candidate objectives. Instead naively trusting every interaction, we propose robots learn conservatively by reasoning real time how relevant correction for test our inference method an experiment with interaction data, demonstrate alleviates unintended in-person user study 7DoF manipulator.