PhysisForcing: PKU Dual-Loss Training Lifts Robot World-Model Closed-Loop Success 50%

Video world models for robot sim routinely hallucinate physically implausible rollouts; PhysisForcing is the first dual-loss training approach (pixel trajectory alignment + semantic relational alignment from a frozen encoder) to push closed-loop robot success from 16% to 24% on WorldArena — a 50% relative gain on the benchmark that feeds directly into embodied-AI policy evaluation.