DiVeR preprint proposes prioritizing decision-critical states in robot verifier training

The researcher says DiVeR learns to rank a robot’s candidate actions from trajectory-level success or failure labels, without step-level annotations or extra environment interaction. They report average success rising from 58.3% to 71.9% across four real-robot tasks compared with single-sample inference. On RoboCasa, they also report over 700× faster verifier inference than a VLM-based baseline.