Yi Xu

Standard robotic reward models routinely assign progress to robots manipulating the completely wrong object, but structuring frozen VLMs into an agentic propose-and-verify inspection loop eliminates these hallucinations without any reward-model training.