The newest motion work treats movement as a compact layer between perception and action. Instead of predicting every future pixel, these papers predict the structure needed to steer a scene, navigate among people, or manipulate an object.

Streaming with scene memory

PhysStream represents sparse velocity increments and a structured scene memory, then allows fine-grained control during generation. The authors report a 33% reduction in FVMD, 12% lower trajectory error, and human preference above 85%. The interesting interface is temporal: an operator can intervene mid-generation without restarting the whole video.

Style as a social forecast

PRISM learns ordinal interaction traits from passive human-human trajectories and uses them for social robot navigation. Its Rank-N-Contrast training and temporal stability score make “interaction style” a predictive signal rather than a label attached after the fact. Simulation results show fewer collisions with small path and time changes. The cautious takeaway is that social motion can be modelled relationally, but evaluation must keep the crowd—not only the robot—in view.

Touch and action in the same world model

DexTouch-WM combines human touch data with bilateral tactile dynamics. A video expert and tactile expert share a world-model backbone; the study uses up to 100 hours of human interaction and five hours of robot supervision. The reported held-out prediction gains point toward a practical route for learning contact-rich manipulation without collecting all demonstrations on the robot itself.

Predict the future that matters

MoWAM replaces future video prediction at inference with an explicit future-motion representation. A mixture transformer samples motion/action candidates and a verifier selects among them. The paper reports stronger out-of-distribution performance and real-task success, supporting a simple engineering intuition: if the downstream decision depends on motion, predict motion directly.

Across these papers, “world model” becomes less synonymous with photorealistic rollout. A useful model may instead remember scene structure, interaction style, tactile consequence, or a small set of candidate futures. That makes its assumptions easier to inspect—and its failures easier to localize.

See our earlier structured-motion digest and contact-focused community notes for the representation layer underneath these predictive systems.

References