PhysisForcing from Peking University and Nvidia Lifts Robot Success From 16% to 24%

Researchers at Peking University and Nvidia introduced PhysisForcing, a training-time framework that adds physics supervision to video diffusion models for robot manipulation. It raised closed-loop manipulation success from 16% to 24% and topped R-Bench, PAI-Bench and EZS-Bench with no added inference cost.

• Closed-loop robot manipulation success rose from 16% to 24%
• Wan2.2-A14B improved by 22.3 points on R-Bench
• Cosmos3-Nano beat Veo 3.1, Hailuo v2 and Seedance 1.5 Pro
• Auxiliary components are discarded after fine-tuning