Simate Unveils Simate-beta, Says Physical AI Model Tops RoboDojo

According to the report, Simate-beta is intended to demonstrate memory, long-horizon task execution, fine manipulation and task adaptation. A company-provided RoboDojo leaderboard updated on Sept. 23, 2026 lists simate-beta first with an average score of 33.95 and an SR of 27.96 percent. The company said the leaderboard result is separate from real-machine demonstrations.

Simate describes its target as zero-shot and few-shot general physical AI operation. The company says a system should not need new data collection, retraining and engineering for every new task in an unfamiliar environment. Its current approach is a general physical fast system that handles real-time perception and rapid action while a slower system handles high-level planning and complex reasoning. Simate says it is scaling the fast system's parameter count to explore emergent zero-shot generalization and to test the limits of deployable edge models.

The company points to two priorities: 4D physical perception, which captures spatial structure and dynamic changes over time, and hierarchical temporal memory, which lets long-horizon tasks remember past events and track states without slowing immediate reactions. Simate says the architecture and model scale will be disclosed in a later technical report. It also says it hopes to coordinate with general reasoning capabilities represented by frontier models such as GPT-6 to advance zero-shot execution of complex tasks.

Behind the model, Simate has organized its research around AI for Physical AI, with AI participating in physical intelligence research and iteration. Its AutoResearch system is designed to automate experiment decomposition, execution and feedback. According to the company, human researchers propose hypotheses, set goals and define constraints, while agents handle configuration changes, training runs, evaluation and result organization.

Simate says the research system has three layers: SiPAI, AutoResearch and AI-native Infra. SiPAI is a pluggable model framework built for automated research, with module boundaries, system interfaces, configuration items and validation processes defined for AI-readable use. The company says it has connected world models, world action models, VLA and VLM architectures. AutoResearch maintains a dynamic research context that includes external papers, internal discussion materials, model code, data ratios, historical experiment logs and the latest evaluation feedback, updating as new evidence arrives. AI-native Infra connects training, inference and evaluation, using task orchestration and resource scheduling to advance dozens of independent research routes in parallel.

According to Simate, all routes first pass through world models and simulation for rapid filtering, and only candidates with potential move to real-machine tests. Problems found on real robots then flow back into the research context, forming a closed loop. The company says this approach helped it reach first place on RoboDojo in three months. Simate has also opened AutoResearch, launched an internal beta and is recruiting participants with 1,000 GPU cards, according to the report. Researchers from MIT, Caltech, Tsinghua University, Peking University and Hong Kong University of Science and Technology have joined the beta, according to the company.

Funding has advanced alongside the research. Simate says it has completed several consecutive financing rounds, each at hundreds of millions of yuan. The report identifies founder Zhang Ying as a former core technical lead at a leading autonomous-driving company, with experience in mapping-based, mapless and end-to-end driving systems and mass production. The company says the driving system he helped build was the closest in China to Tesla FSD. Other team members named in the report include Zhan Fangneng, an assistant professor at HKUST and head of World Mind Lab, who researches world models and physical AI, and Ji Ma Zeyu, a young scientist working on 3D perception, dexterous manipulation and humanoid whole-body control. Before joining Simate, Ji was a founding member of Silicon Valley's Assured Robot Intelligence, which was later acquired by Meta.

Simate defines its goal as Physical RSI and describes three categories. Weak SI has clear boundaries and fast closed loops, with researchers setting goals and constraints while handling exceptions; the RoboDojo result is described as first-stage engineering validation. Medium SI has a clear direction but unknown answers, with agents organizing context, implementing proposals, running experiments and comparing results, while researchers select hypotheses, interpret conflicting evidence and make key trade-offs. Strong SI involves discovering research questions themselves, including improving generalization to entirely new tasks; the company says this still requires senior researchers, with agents supporting literature, evidence, experiments and route exploration. Simate says the three categories are not hierarchical and can coexist. Human-in-the-loop remains, and researchers still control key directional judgments, according to the company.

Simate says it has formed a technical approach for zero-shot generalization to complex tasks and plans to release stage results by the end of this year. Three research efforts, including the model and automated research, will be published through papers and technical reports, with results to be open-sourced in stages. The report lists the AutoResearch platform website as mate-robot.cn.

Editor's Summary Simate's Simate-beta ranked first on a company-provided RoboDojo leaderboard, according to QbitAI, while the company advances an AI-native research system and plans year-end results. The startup has also raised multiple hundreds-of-millions-yuan rounds and opened AutoResearch to external beta testers.