QiYuan Embodied Research Center Releases PrimeSim World Model Simulator, Tops Two Robot Authority Benchmarks

On November 4, the Qiyuan Embodied Research Center released the world model simulator PrimeSim. This product is based on a pre-trained video generation model, modified through two stages of mid-training and post-training, and possesses interactive reasoning capabilities to replace traditional simulators, which can be used to conduct policy training for robot manipulation tasks. Related weights, inference code, and training code will be open-sourced on November 6.

Simulation environments are an important foundation for training robot control policies, but relying on manually constructed simulation scenes makes it difficult to support large-scale deployment. According to the Qiyuan Embodied Research Center, an ideal robot simulator needs to simultaneously satisfy requirements such as parallel scheduling of tens of thousands of trajectories, precise scene resetting without hardware wear, reducing engineering customization across scenes, and narrowing the gap with the real environment. The team believes that relying solely on video generation models to achieve “visual realism” is still insufficient; as a simulation base, it also needs to strictly obey action instructions from policy outputs, maintain geometric and semantic consistency between multi-camera views, and maintain state coherence over time spans far longer than the training context window.

In authoritative benchmark evaluations, PrimeSim performs outstandingly. On the EWMBench benchmark, the simulator ranks first with a comprehensive score of 6.09, and sub-metrics HSD (1.000), nDTW (1.000), BLEU (0.570), CLIP (0.954), and Logics (1.000) all achieved first place. In five evaluations of WorldSimProbe based on LIBERO, PrimeSim topped three dimensions: global trajectory coverage (67.9), action source behavior maintenance (72.3), and interaction dynamics (37.1).

According to the research team, PrimeSim adopts a two-stage modification paradigm, gradually transforming a native bidirectional video generation model into an interactive, autoregressive rolling world model simulator. In the middle training stage, the model is based on 30110 hours of selected robot demonstration data, covering 22 types of heterogeneous robot platforms including humanoid, dual-arm, single-arm, and handheld UMI devices, where humanoid robot samples account for 54.8% and real-world data accounts for 84%, and it achieves cross-embodiment compatibility and strong binding of control signals around three core designs: unified action representation, frame-aligned dual-path conditional injection, and viewpoint-aware interleaved attention. In the post-training stage, through three mechanisms—high-speed generation links, self-conditional stability, and hierarchical temporal memory—the model can still stably infer even after exceeding the fixed training context window.

Experimental data show that, in the RoboTwin 2.0 hold-out set and a single A100 environment, PrimeSim, through consistency distillation and DMD superposition, compresses the denoising steps to 4 steps, and the speed is improved from 1.12 FPS of the 50-step denoising baseline to 10.04 FPS, an overall speedup of about 9 times.

As a policy simulation base, PrimeSim demonstrates high alignment with traditional simulators. The experiment selected 8 OpenVLA-OFT policies, and under 300 shared initial state settings, the correlation coefficient of the policy success rate between PrimeSim and the native LIBERO simulator reached 0.99, with an average absolute error of 2.8. The team also directly conducted GRPO reinforcement learning on the policies within PrimeSim, and the results showed that the success rate of the policies on the native simulator increased by 11.5%, and the evaluation metrics within PrimeSim rose by 11.8%, proving that the policy gains obtained in this neural simulator can be effectively transferred to traditional physical simulation environments.

In terms of few-shot embodiment transfer, the training endows the model with strong cross-embodiment capabilities: adapting to xArm 7, ViperX 300, and UMI grippers requires only 200, 100, and 200 gradient updates, respectively. In the AgileX Piper scenario, which was not seen in the training set, after approximately 200 steps of fine-tuning on a single A100 card, the cloth deformation can follow motion instructions, and multi-view states remain consistent. In terms of inference performance, under a configuration of a single A100-80GB, three views, and a single image resolution of 288×384, PrimeSim can achieve 10.12 TempFPS and 30.35 ViewFPS, with a throughput capacity approximately 4.5 times that of GE-Sim 2.0 and 11.6 times that of Ctrl-World.

The Qiyuan Embodied Research Center stated that this release validates the feasibility of transforming a general video generation model into a robot world model simulator through a “mid-training + post-training” progressive pipeline, and the relevant results balance action fidelity, long-sequence inference capability, and engineering practicality, providing a new path for the scaling of robot control policy training.

 (LeFeng Net (Official Account: LeFeng Net)）