PhysisForcing: Physics-Reinforced World Models Improve Robot Manipulation Success by 50%
Peking University / NVIDIA
PhysisForcing applies hierarchical physics supervision to video-generation-based world models for robot training: pixel-level trajectory alignment using reference point trajectories, and semantic-level relational alignment from a frozen video encoder. Improves closed-loop manipulation success from 16.0% to 24.0% and achieves 3.7–22.3% gains over baselines. Model-agnostic approach demonstrated on Cosmos3-Nano.
Why it matters
Physical plausibility of world models is a key bottleneck for sim-to-real transfer in robotics. 42 upvotes on HuggingFace Daily Papers on June 29, 2026.
Importance: 2/5
42 HF Daily upvotes; 50% relative improvement in robot manipulation success via physics-enforced world model training