💡 The paper proposes PhysisForcing, a framework for enhancing the physical consistency of embodied video generation models for robotic manipulation. The problem with existing video generation models is that they can produce physically implausible manipulations, such as discontinuous motion trajectories and inconsistent robot-object interactions. This is mainly due to the deformation of moving objects and implausible spatio-temporal correlations among interacting entities, particularly during contact.

To address this issue, PhysisForcing uses a scalable training framework that focuses supervision on physics-informative regions through joint optimization of pixel-level and semantic-level features. The framework consists of two losses: a pixel-level trajectory alignment loss that supervises features using reference point trajectories, and a semantic-level relational alignment loss that aligns features with inter-region relations extracted from a frozen video understanding encoder.

The method is evaluated on several benchmarks, including R-Bench, PAI-Bench, and EZS-Bench, and the results show that PhysisForcing consistently improves embodied video generation over strong baselines. Specifically, it improves the Wan2.2-I2V-A14B and Cosmos3-Nano base models on R-Bench by 22.3% and 9.2%, respectively, with the Cosmos3-Nano variant attaining the best overall score.

Furthermore, the paper demonstrates that PhysisForcing can be used as a world model under the WorldArena action-planner protocol, which raises the closed-loop success rate from 16.0% to 24.0% and further improves downstream policy success. This indicates that physically aligned video models yield stronger representations for robotic manipulation. Overall, the paper contributes a novel framework for enhancing the physical consistency of embodied video generation models, which has the potential to improve the reliability and performance of robotic manipulation systems.