PrimeSim: A Generalist World Model Simulator from Large-Scale Cross-Embodiment Demonstrations

PrimeSim. A Generalist World Model Simulator from Large-Scale Cross-Embodiment Demonstrations. By Travis Davies and twelve others, from PrimeBot Research, Southeast University, and six other institutions. Published on the third of November.

A robot world model that is meant to serve as a simulator has a harder job than generating plausible video. It has to follow the actions it is commanded to take, faithfully. It has to stay consistent across several camera views. And it has to keep long rollouts coherent, instead of drifting as it goes. A video that merely looks right fails all three tests if the arm does not move the way it was told, if two cameras disagree about the same scene, or if the scene falls apart a few seconds in. A simulator is only useful when what it shows can be trusted as the consequence of the actions fed into it.

PrimeSim is an action-conditioned, multiview world model simulator built to meet those three demands. It gets there in stages, through progressive midtraining and then posttraining, starting from a pretrained video generator. The midtraining data is large: thirty thousand one hundred and ten hours of robot demonstrations, spanning twenty-two platforms.

Each demand gets its own component. For action following, there is a unified action representation shared across embodiments, plus frame-aligned dual-path conditioning that strengthens the correspondence between actions and frames. For multiview consistency, there is view-aware interleaved attention with explicit camera identities. It lets information pass between views while keeping the representations specific to each view.

Long horizons are handled in posttraining. A framework the authors call fast video, slow memory turns the bidirectional video generator into an efficient autoregressive simulator. It combines fast causal generation, self-rollout training and hierarchical temporal memory. Together these reduce the error that accumulates when the model conditions on histories it generated itself, which is what allows coherent simulation over long horizons. The split in the name describes the division of labour: fast generation of the video itself, and a slower, hierarchical memory that carries what happened earlier.

PrimeSim runs at ten point one frames per second. It stays stable over long rollouts on manipulation tasks with both rigid and deformable objects. On platforms held out from training, it adapts within one hundred to two hundred gradient steps, where the pretrained backbone does not. Against five open-source action-conditioned world models, it leads four of the five WorldSimProbe suites and posts the best overall EWMBench score. The authors are candid that fine-grained local calibration still trails the field. Its rollouts are also faithful enough to support policy optimization based on GRPO, which is the use a simulator is ultimately for: training a policy inside it.

The model weights are released. What PrimeSim argues for is a simulator assembled from a video generator by targeting each failure separately: an action representation and conditioning for following commands, view-aware attention for consistency across cameras, and a fast-video, slow-memory posttraining for staying coherent over time.