Simate's General Physical Fast System Model Tops RoboDojo Benchmark
Physical intelligence company Simate disclosed on September 24 that its general physical fast system model, developed with the AI auto-research system AutoResearch, ranked first on the RoboDojo benchmark for physical AI. Real-robot demonstrations showed the model handling complex long-horizon tasks, retaining memory, and performing high-precision manipulations.
RoboDojo is a benchmark for general robot manipulation, with evaluation dimensions covering generalization, memory, fine manipulation, long-horizon task execution and open tasks. According to Simate, the general physical fast system integrates a 4D physical perception and memory mechanism that understands the physical environment through depth, geometry, motion and contact relations, while organizing historical information around task progress, continuous state tracking and immediate feedback, as an attempt to balance long-horizon consistency with real-time response. In the demo, the model used historical information to continuously execute a multi-step tea-making task.
Founded about three months ago, Simate has completed several hundred million yuan in cumulative financing and launched its official website. Its core team includes technical talent from autonomous driving end-to-end mass production, world model, and humanoid robot control backgrounds.
Simate has proposed a Physical RSI (Physical Recursive Self-Improvement) technical route that links model R&D, data processing, experiment evaluation and real-robot deployment, letting AI agents take on more research work and using accumulated experience from each cycle to improve models and research methods. The company said AutoResearch has been opened for trial use and is already being used by researchers at Tsinghua, MIT and HKUST.
