PrimeSim Cross-Embodiment World Model Simulator Released

The research team released PrimeSim, an action-conditioned multi-view world model simulator, and made public the model weights, inference code, and training code. The model was mid-trained on robot demonstration data covering 22 platforms and 30,110 hours, runs at 10.1 FPS, remains stable in long-horizon rollouts on rigid and deformable manipulation tasks, and can transfer within 100 to 200 gradient steps to held-out platforms to which the pretrained backbone cannot transfer. The team states that in comparisons with five open-source action-conditioned world models, PrimeSim leads on four of the five WorldSimProbe test suites and has the highest overall EWMBench score, but still lags behind comparable models in fine-grained local calibration; its rollouts are sufficient to support GRPO-based policy optimization.

AI 根据报道生成 · 7 days ago更新