PrimeSim world model simulator from large-scale cross-embodiment demonstrations
Original title: PrimeSim: A Generalist World Model Simulator from Large-Scale Cross-Embodiment Demonstrations
PrimeSim is an action-conditioned multiview world model simulator that runs at 10.1 FPS and transfers to held-out platforms after 100 to 200 gradient steps.
PrimeSim starts from a pretrained video generator and is midtrained on 30,110 hours of robot demonstrations from 22 platforms. It uses a unified cross-embodiment action representation and frame-aligned dual-path conditioning to improve action following. View-aware interleaved attention with explicit camera identities maintains multiview consistency. A fast-video slow-memory posttraining framework enables efficient autoregressive simulation with self-rollout training and hierarchical temporal memory to support long-horizon rollouts.
- Stable over long-horizon rollouts on rigid and deformable manipulation tasks
- Leads four of five WorldSimProbe suites
- Best overall EWMBench score
- Faithful rollouts support GRPO-based policy optimization
- Transfers to held-out platforms within 100-200 gradient steps
From the original article
1 min read at huggingface.co
A robot world model that serves as a simulator must do more than generate plausible videos: it must faithfully follow commanded actions, maintain consistency across camera views, and sustain coherent long-horizon rollouts. We present PrimeSim, an action-conditioned, multiview world model simulator that addresses these challenges through progressive midtraining and posttraining.
By Travis Davies, Wentao Ma, Lihao Guo · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.