UPenn's Mao Jiayuan details RAPID and a TR metric that lifts real-robot generalization to 83%

Mao Jiayuan of UPenn, MIT and Stanford presented the RAPID framework at IROS 2026, which learns non-grasping manipulation from a single human demonstration.

She also introduced the TR metric, which measures whether a video-action model's action head actually uses predicted future frames, and says adaptive guidance based on TR raised dual-arm real-robot generalization on complex composite tasks to 83%.