MotorMind: Scaffolding General Vision Language Models for Zero-Shot Robot Manipulation
MotorMind lets a general-purpose VLM drive robot manipulation through mid-level actions and feedback, without task-specific policy training.
The paper reports 66.7% success on base LIBERO-PRO suites and 53.8% under perturbations. Prior zero-shot methods evaluated in the paper topped out at 13.3% and 19.2% on those settings. On a real xArm6 robot, the same interface averaged 95% success across direct manipulation and human-perturbation tests. The authors say remaining failures mainly involve visual grounding, embodied reasoning, and action knowledge. HF Daily Papers' note
The paper reports 66.7% success on base LIBERO-PRO suites and 53.8% under perturbations. Prior zero-shot methods evaluated in the paper topped out at 13.3% and 19.2% on those settings. On a real xArm6 robot, the same interface averaged 95% success across direct manipulation and human-perturbation tests. The authors say remaining failures mainly involve visual grounding, embodied reasoning, and action knowledge. HF Daily Papers' note
score 5