Graph-Reweighted Behavior Cloning with Human Interventions for Precise Robotic Manipulation

This research proposes Graph-Reweighted Behavior Cloning (GR-BC), an offline iterative post-training framework for precise robotic manipulation. GR-BC builds a task-conditioned goal-reachability graph over demonstrations and human-intervention-assisted rollouts, using reverse graph search to compute minimal steps to task success. The method weights action chunks based on their impact on task completion, reusing recoverable segments from failed trajectories and down-weighting regressive actions. GR-BC improves single-round supervised fine-tuning and achieves 91% success without trajectory comparisons or progress labels.