PrimeSim: A Generalist World Model Simulator from Large-Scale Cross-Embodiment Demonstrations
PrimeSim reports 10.1 FPS robot simulation and adaptation to unseen platforms within 100–200 gradient steps.
The researchers trained a pretrained video generator on 30,110 hours of robot demonstrations spanning 22 platforms. Its action conditioning, camera-aware attention and temporal memory aim to keep commanded movements, multiple views and long simulations consistent. Against five open-source action-conditioned world models, it leads four of five WorldSimProbe suites and achieves the best overall EWMBench score, though fine-grained local calibration still trails competitors. The authors report releasing model weights, inference code and training code.
HF Daily Papers' note
The researchers trained a pretrained video generator on 30,110 hours of robot demonstrations spanning 22 platforms. Its action conditioning, camera-aware attention and temporal memory aim to keep commanded movements, multiple views and long simulations consistent. Against five open-source action-conditioned world models, it leads four of five WorldSimProbe suites and achieves the best overall EWMBench score, though fine-grained local calibration still trails competitors. The authors report releasing model weights, inference code and training code.
HF Daily Papers' note
score 5