PrimeSim robot world model hits 10.1 FPS, leads 4 of 5 probes

PrimeSim, a multiview robot world model from PrimeBot Research with Shanghai Jiao Tong University, The Chinese University of Hong Kong, Peking University and five other labs, runs at 10.1 FPS on a single A100 and leads four of the five WorldSimProbe suites against five open-source action-conditioned baselines. The paper posted on Hugging Face reports an overall EWMBench score of 6.09, ahead of GE-Sim (LTX) at 5.82.

The scale behind that result is unusual for an open release. Midtraining pulled from 30,110 hours of robot demonstrations spanning 22 platforms, 20 dataset families and 1.52 million episodes, with roughly an 84:16 real-to-simulated ratio. Few-shot adaptation to held-out embodiments like the xArm 7 and ViperX 300 converges within 100 to 200 gradient steps, where the pretrained video backbone, per the paper, reaches no action-following rollout inside a 2,000-step budget.

The authors flag the weak axis directly. The abstract notes that "fine-grained local calibration still trails the field," even as the model takes the lead on action realisation and interaction dynamics. Interaction Dynamics (T5) lands at 37.1, a 48.4% relative gain over the strongest prior entry but still a low absolute score.

Where the paper reaches furthest is using the simulator as an RL training ground. A 200-step GRPO run inside PrimeSim lifts an SFT policy from 61.7% to 73.2% in-simulator success, and PrimeSim's own policy-success estimate tracks the real simulator within 1.6 percentage points. The team releases weights, inference code and training code.