One World Model, Many Robot Bodies: The New Playbook for Robot Learning

Shared actions across bodies

In a preprint published October 28, researchers introduce RynnWorld-Latent, a cross-embodiment video world model that uses 3D-grounded latent actions to unify heterogeneous interaction data. The model was pretrained on 23 datasets comprising approximately 553.5 million frames, and it uses frozen RynnLAM descriptors as a shared conditioning interface. The framework supports digital teleoperation to generate action-video demonstrations and latent-action synthesis to transfer source motion to an edited initial frame, and the model is adapted to target robots through post-training. Synthetic mid-training improved policy success rates on LIBERO and LIBERO-Plus, and both generation routes improved real-world policy success rates over real-demonstration-only training. m:…3517

The idea behind the shared interface is worth stating plainly. Different robots have different joints, grippers, and sensors, so data collected on one body does not transfer naively to another. By conditioning on latent actions grounded in three-dimensional space rather than on robot-specific commands, the preprint argues, one world model can learn dynamics that are useful across bodies and then be adapted to a specific target robot. That is a data-efficiency claim as much as a modelling claim: if synthetic data from a shared world model improves real-world policy success, the bottleneck for a new robot shifts from collecting demonstrations on that exact hardware to adapting a model that already understands how the world moves. m:…3517

Predict physics, not video

In a second preprint, researchers introduce PDRVLA, a vision-language-action model that improves robot action generation by integrating a lightweight physical dynamics module. The module predicts future visual features, including RGB, depth, and optical flow, to condition action generation, and it uses auxiliary supervision from a frozen pretrained video world model. PDRVLA avoids online video generation latency, a practical advantage for a robot that must act in real time rather than wait for a video to render. Training data included human-to-robot synthesis, subtask annotation, and embodied chain-of-thought annotation. m:…7626

The benchmark numbers reported in the preprint are strong. PDRVLA achieved a 99.3 percent success rate on the LIBERO benchmark, 89.9 percent on LIBERO-Plus, 93.6 percent on RoboTwin 2.0, and 81.5 percent on SimplerEnv, and ablations showed improved future prediction with world-model supervision. Those numbers are benchmark results in a preprint, not evidence of deployment on a factory floor, but they support the design bet: a model that predicts the physical features that matter for the next action can outperform a model that spends its compute generating full video it does not need. m:…7626

Worlds that check themselves

In a third preprint, researchers introduce PhysEvolve, a framework in which vision-language models learn causal physical reasoning without human data. A shared vision-language model proposes executable world programs and evidence schemas at each round, a deterministic physics engine freezes each valid world, and exact verifier feedback from task solving updates the shared policy. The system synthesises discrete, answer-bound visual tasks, and the policy update uses GRPO, with Qwen3-VL-8B-Instruct as the base model. PhysBench Throwing performance improved by 8.8 points, ContPhy Counterfactual Removal by 7.4 points, and ContPhy Goal-Pit by 24.4 points, with gains concentrating on interaction-sensitive and altered-world questions. m:…1942

PhysEvolve matters because it removes a label bottleneck. If a physics engine can validate a generated world, the system can create its own training signal instead of waiting for humans to annotate physical reasoning, and the preprint reports zero-shot transfer of causal principles to unseen complex dynamics. That is a different kind of scaling from collecting more robot demonstrations: scaling the worlds themselves, with a verifier that rejects worlds that do not obey the physics it encodes. The risk, which the preprint does not resolve, is that the engine encodes only the physics its authors implemented, so transfer beyond that envelope remains an open question. m:…1942

Cheaper foresight

In a fourth preprint, researchers introduce the Just Action Model, which decouples backbone initialisation, future observation objectives, and inference machinery. The foresight loss trains current-observation features to predict frozen encodings of the recorded future, which eliminates the need for visual generation during training or inference. Just Action Model policies outperformed matched FastWAM and ImageWAM counterparts on LIBERO and LIBERO-Plus, and the model reduced estimated online training FLOPs per sample per update by 73 percent relative to FastWAM. The model supports generator, vision-language, and self-supervised backbones. m:…9642

Together, the four preprints describe a converging playbook. Share action representations across bodies, predict only the future features that actions need, generate training worlds that a verifier can check, and keep the foresight loss cheap enough to train at scale. Each paper attacks a different cost in robot learning: data collection, inference latency, human labelling, and training compute. None of them claims a general-purpose robot, and all of them are preprints whose benchmark gains still need to survive contact with varied hardware, but the direction of travel is consistent and the cost reductions are large enough to change what a robotics team can afford to try. m:…3517 m:…7626 m:…1942 m:…9642