World Models — Physical Prediction and Transfer

World-model-based control generalises across an unusually broad benchmark suite under one algorithmic configuration, establishing that learned latent dynamics can support action across heterogeneous task domains. Hafner et al. report DreamerV3 mastering more than 150 diverse control tasks with a single configuration and outperforming specialised methods across the evaluated benchmarks. This is substantive evidence that world-model planning can support broad control capability, but the evidence is primarily benchmark and simulated-environment generality rather than the record's harder requirement of reliable physical transfer across materially changed real-world objects, environments or embodiments.

This is the strongest current supportive evidence for bounded physical transfer. Assran et al. pre-train V-JEPA 2 on more than one million hours of video and then post-train an action-conditioned world model using less than 62 hours of unlabeled DROID robot video. V-JEPA 2-AC is deployed zero-shot on Franka arms in two different laboratories for image-goal reaching, grasping and pick-and-place without collecting robot data in the target environments and without task-specific training or reward. The result demonstrates useful action-conditioned prediction surviving a meaningful environment shift. It does not yet establish broad task transfer or cross-embodiment transfer because the deployment remains within a constrained manipulation setting and the same general robot platform family.

Interactive world simulation advances materially, but the result is not equivalent to reliable physical transfer. Google DeepMind reports that Genie 3 generates 720p interactive environments at roughly 24 frames per second, responds to user actions and maintains environmental consistency for several minutes. DeepMind also states important limits: agent action space is constrained, multi-agent interaction remains difficult, real-world locations are not represented with perfect accuracy, and continuous interaction lasts minutes rather than hours. The evidence therefore supports increasingly capable action-responsive simulation while leaving the record's physical-fidelity and real-world functional-utility commitments unresolved.

Direct benchmark evidence contests the assumption that realistic generated futures are reliable enough for robot planning. Yang et al. introduce MiraBench with more than 16,000 human judgments across 12 representative world-model configurations, evaluating physics adherence, action-following fidelity and optimism bias under failure-inducing actions. The study reports that visual fidelity is a poor proxy for action fidelity, increasing model scale does not reliably improve action following, and optimism bias is pervasive. This directly pressures intervention fidelity and physical fidelity: a model can generate a convincing future while failing to represent what the commanded action would actually cause.

State-of-the-art video world models remain brittle when a physically meaningful variable is changed while the rest of a scene is held stable. Cai et al. construct 319 paired interventions from nuScenes and DROID and evaluate whether generated futures diverge in the way the changed physical condition requires. Across nine models, no system exceeds 52% on the paired score and every tested model fails a substantial fraction of causal interventions. The result is important because each individual video may look plausible while the pair reveals that the model did not track the causal consequence of the intervention. This directly contests the claim's intervention-fidelity commitment.

Embodiment-grounded evaluation shows that visually plausible predicted behaviour can still fail when converted into action and tested for executability in physically grounded simulation / real-to-sim environments. Jiang et al. convert generated human-hand and robotic manipulation videos into embodied action sequences and evaluate those behaviours in simulation-based execution environments rather than establishing physical-robot validation. Across evaluated world models, reliably generating executable behaviour remains an open challenge; reported failure modes include spatial-reasoning errors, unstable contact prediction and non-physical deformation. Fine-tuning on manipulation data improves results but does not remove the physical inconsistencies. This contests physical fidelity and functional utility while remaining bounded to simulation-based embodied executability.

Mechanistic analysis supplies supportive context for the possibility that useful physical prediction need not resemble an explicit physics engine. Joseph et al. probe large-scale video encoders and report an intermediate-depth transition where speed, acceleration and motion-direction information becomes accessible, with physical variables represented in distributed high-dimensional structures rather than compact factorised state variables. The result bears on mechanism, not transfer: it supports the plausibility that learned representations contain action-relevant physical structure while not establishing that the structure remains sufficiently faithful for reliable planning under changed environments or embodiments.

Distributional evaluation adds a distinct reliability failure mode beyond visual plausibility or single-trajectory physical error. CaliBench (arXiv:2608.16829) evaluates six image-to-video models across nine controlled physical scenes with analytically known stochastic outcome distributions, using repeated generations to test whether model output frequencies match those distributions. The study finds widespread significant miscalibration and severe cases of probability-mass concentration or mode collapse, showing that a generated future can look physically plausible while the model represents the range and likelihood of possible physical outcomes incorrectly. This directly contests the record's physical-reliability commitment because planning under uncertainty depends not only on plausible trajectories but on sufficiently faithful outcome probabilities. The evidential boundary is strict: CaliBench tests image-to-video stochastic calibration in controlled scenes, not action-conditioned robot planning, cross-embodiment transfer or successful physical execution, so it adds contesting mass within the existing reliability shape rather than establishing failure of the full transfer claim.

Li et al. collect human manipulation trajectories in a purpose-built VR cooking game, extract contact-point and sub-goal affordance cues, pre-train an affordance model on those game-collected data, and adapt downstream robot policies with a small number of real-robot demonstrations. The authors report average success-rate gains of 10.0 percentage points in simulation and 18.3 points on real robots in the few-shot setting against policies without that game-data pretraining. This is a controlled, task-specific supportive result for transferring selected manipulation knowledge from a virtual environment to physical action. It is not evidence that an unrestricted video-game corpus yields a generally reliable action-conditioned world model: the game was designed for data collection, the cues were explicitly extracted, and real-robot adaptation remained necessary. The result is a September 2026 preprint; code and platform release is promised upon acceptance, so independent replication and broader task or embodiment transfer remain open.