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

Researchers present PrimeSim, an action-conditioned, multiview world model simulator designed to address challenges in robot simulation like video generation, action following, and long-horizon consistency. The model utilizes progressive midtraining on robot demonstrations, unified cross-embodiment action representations, and a fast-video, slow-memory posttraining framework. PrimeSim runs at 10.1 FPS, achieves strong scores across benchmark suites, and transfers to held-out platforms within 100 to 200 gradient steps. Open article →