The paradigm of "AI researching AI" has officially entered the physical world. Just three months after its founding, #physical AI startup Simate (Silicon Mate) unveiled its first universal physical fast system, Simate-beta, demonstrating advanced long-horizon task execution, fine manipulation, and adaptability. The model grabbed the top spot on the RoboDojo leaderboard without any benchmark-specific optimization.
Simate's founding team previously scaled end-to-end autonomous driving models to match Tesla FSD performance and achieved mass-production deployment. Distinct from traditional robotic startups, Simate’s core philosophy is "AI for Physical AI," enabling AI agents to drive the research and iteration of physical intelligence. They built an AI-native research system powered by the AutoResearch platform and Sinfra infrastructure. Researchers from MIT, Caltech, Tsinghua, and Peking University have joined the closed beta, backed by hundreds of millions of RMB in funding.
Simate solves the task-specific adaptation bottleneck of #embodied AI through its universal "Fast System" (or System 1). While a "Slow System" (System 2, such as future GPT-6 models) handles high-level reasoning, the fast system processes real-time perception and milli-second level motor control. Simate-beta implements 4D physical perception to capture spat-temporal dynamics and utilizes hierarchical temporal memory to track past states without lagging real-time execution.
Powering this rapid iteration is a three-tier architecture consisting of SiPAI (a pluggable model framework), AutoResearch, and AI-native Infra. By making the pipeline AI-researchable, the AutoResearch engine autonomously executes and evaluates experiments. On the September 23, 2026 update of the #RoboDojo benchmark, Simate-beta achieved a first-place score of 33.95 with a success rate of 27.96%. The team plans to release zero-shot generalization results by the end of the year, steering physical AI toward its own "GPT-3 moment."
[AgentUpdate Depth Analysis] Simate's "AI for Physical AI" paradigm represents a significant breakthrough for the Embodied AI and AI Agent ecosystems. While digital-native agents have rapidly matured, physical agents have long struggled with high simulation-to-real transfer gaps and slow R&D cycles. By structuring model frameworks to be fully machine-interpretable via SiPAI and AutoResearch, Simate transitions robotic engineering from human-in-the-loop trial to automated agentic optimization. This System 1 and System 2 decoupling ensures that low-latency physical feedback loop operates independently from slow cognitive chains. Simate’s approach offers a highly inspiring blueprint: the future of Embodied AI relies not just on scaling model size, but on building automated, self-evolving research flywheels that seamlessly connect digital intelligence with physical constraints.
