Counterfactual Behavior Cloning: Offline Imitation Learning from Imperfect Human Demonstrations
Research
Source: arXiv cs.ROPublish time unverified
arXiv:2505.10760v2 Announce Type: replace Abstract: Learning from humans is challenging because people are imperfect teachers. When everyday humans show the robot a new task they want it to perform, humans inevitably make errors (e.g., inputting noisy actions) and provide suboptimal examples (e.g., overshooting the goal).