GR-BC learns robot skills from failed attempts, hitting ~91% success rate

jiqizhixin · x · 2026-11-18

Researchers from Zhejiang University, TeleAI, and Gamma Robotics introduce GR-BC, which learns from demonstrations, robot rollouts, and human corrections. It builds a goal-reachability graph across attempts (multi-view + gripper consistency), uses shortest-path search to estimate remaining steps, and weights action chunks in behavior cloning so useful fragments of failed trajectories count. With X-VLA it reports 91% mean success across four real manipulation tasks.

Original post →

More from Embodied

Embodied channel →