Robot training preprint reports 91% success by reweighting imperfect demonstrations

Researchers report that Graph-Reweighted Behavior Cloning reached 91% success after a few offline training iterations across four precise manipulation tasks and three vision-language-action models, which translate visual and language inputs into actions. The preprint describes a method for learning from useful portions of both successful and failed robot trajectories. The method builds a graph of recorded actions and uses reverse search to estimate the remaining steps to success. It gives greater training weight to actions that advance the task and uses human interventions as recovery paths. The reported results involve intervention-assisted data and do not establish performance beyond the tested tasks and models.