PhysisForcing: Physics Reinforced World Simulator for Robotic Manipulation

TL;DR

PhysisForcing is a finetuning framework that targets the physical-instability failure mode of video world simulators used for robotic manipulation — discontinuous motion trajectories and inconsistent robot-object interactions during contact. Through an empirical attribution it identifies two dominant causes (deformation of moving objects, implausible spatio-temporal correlations among interacting entities) and applies two physics-informed auxiliary losses on top of standard finetuning: a pixel-level trajectory-alignment loss that supervises DiT features against reference point trajectories, and a semantic-level relational-alignment loss that aligns DiT features to inter-region relations from a frozen video-understanding encoder. On the R-Bench / PAI-Bench / EZS-Bench robot-oriented suite it lifts Wan2.2-I2V-A14B by 22.3% and Cosmos3-Nano by 9.2% over base (7.1% / 3.7% over vanilla finetuning), and under the WorldArena action-planner protocol the closed-loop success rate climbs from 16.0% to 24.0%.

Key claims

Method

PhysisForcing leaves the base DiT architecture untouched and adds two auxiliary losses during finetuning, each motivated by one of the diagnosed instability modes [Abstract]. The pixel-level trajectory alignment loss takes reference point trajectories (sampled on physics-informative regions of the scene) and supervises intermediate DiT features so that they remain consistent with how those points should move — directly attacking the “deformation of moving objects” failure. The semantic-level relational alignment loss runs a frozen video-understanding encoder over the target clip to extract inter-region relations and aligns the DiT’s features with those relations — attacking the “implausible spatio-temporal correlations among interacting entities” failure, particularly during contact. The two losses are jointly optimized with the standard denoising objective. Notably both auxiliary signals are concentrated on physics-informative regions rather than applied uniformly, which is the paper’s structural bet: don’t supervise pixels equally, supervise contact and trajectory.

The framework is base-model-agnostic — the two models exercised in the paper, Wan2.2-I2V-A14B and Cosmos3-Nano, are very different (MoE video DiT vs. NVIDIA’s omnimodal physical-AI model), and both improve, suggesting the auxiliary losses are not co-adapted to a specific backbone [Abstract].

Results

On R-Bench (the robot-oriented video-generation benchmark from Rethinking Video Generation Model for the Embodied World (ReVidGen / RBench / RoVid-X)), PhysisForcing improves Wan2.2-I2V-A14B by 22.3% and Cosmos3-Nano by 9.2% over the base models [Abstract].
Against vanilla finetuning on the same data, the gains are 7.1% (Wan2.2-I2V-A14B) and 3.7% (Cosmos3-Nano) on R-Bench — i.e., roughly a third of the total improvement is attributable specifically to the physics-informed losses rather than to additional data exposure [Abstract].
The Cosmos3-Nano variant attains the best overall score on the combined R-Bench / PAI-Bench: A Comprehensive Benchmark For Physical AI / EZS-Bench evaluation [Abstract].
Under the WorldArena action-planner protocol, plugging PhysisForcing in as the world model raises closed-loop success rate from 16.0% → 24.0% [Abstract].
Downstream policy success also improves when using PhysisForcing-generated rollouts as a representation substrate [Abstract].
The paper reports no specific per-axis breakdowns in the abstract; sub-metric deltas (structural consistency / physical plausibility / action completeness from R-Bench, and the PAI-Bench tracks) are pending full-paper fetch.

Why it’s interesting

PhysisForcing is the first filed paper that intervenes on the physics-instability axis that the R-Bench paper (Rethinking Video Generation Model for the Embodied World (ReVidGen / RBench / RoVid-X)) diagnosed — and the author overlap (Yufan Deng, Zilin Pan, Daquan Zhou) makes it explicit that this is the same group following their own diagnostic finding (“significant deficiencies in generating physically realistic robot behaviors”) with a training-time fix. The two physics-informative auxiliary losses also give the cluster a concrete recipe for the long-standing complaint, raised by Towards World Simulator: Crafting Physical Commonsense-Based Benchmark for Video Generation (PhyGenBench) and VideoPhy-2: A Challenging Action-Centric Physical Commonsense Evaluation in Video Generation, that scale alone does not produce physical commonsense — here, targeted representation alignment at contact regions does much of the work that data scale doesn’t. As a piece of World Foundation Models research it joins the language-conditioned-rollout line of Qwen-RobotWorld Technical Report: Unifying Embodied World Modeling through Language-Conditioned Video Generation and Cosmos 3: Omnimodal World Models for Physical AI but argues that the right axis for improvement is not just architecture (dual-stream MMDiT) or data scale, but per-region physics-aware supervision. It also lands a non-trivial closed-loop number (16% → 24% on WorldArena) — concrete evidence that better physical plausibility in the world model translates to better downstream policy success, which is the empirical pillar Direct Video-Action Models — Causal Video Models Are Data-Efficient Robot Policy Learners and ImageWAM: Do World Action Models Really Need Video Generation, or Just Image Editing? have been building toward.