On September 24, physical intelligence startup Simate released a progress update covering its first three months of operation.

According to the company, Simate used AutoResearch, its automated AI research system, to help develop a general-purpose physical fast-system model. That model has taken first place on RoboDojo, a physical AI benchmark, and has demonstrated long-horizon task execution, environmental memory and high-precision manipulation on real robots.

Simate has raised several hundred million yuan to date, with team members drawn from autonomous driving, world models and humanoid robot control. The company is pursuing a new technical route: putting AI itself to work on the next generation of robot intelligence.

From Autonomous Driving to Physical Intelligence

Simate founder Zhang Ying previously served as a technology lead at a leading autonomous driving company, with a long track record in end-to-end autonomous driving and complex intelligent systems.

Zhang sees autonomous driving and physical intelligence as resting on similar foundations. Both types of system have to perceive real environments, interpret external change, make decisions and continuously adjust their actions.

Robots, though, face harder tasks. Take grasping: a change in an object's shape, material or placement can alter the outcome. A robot has to understand its environment, execute fine motor control, and handle anomalies when they arise.

To advance its Physical RSI approach, Simate has assembled researchers from both academia and industry.

Team member Zhan Fangneng is an assistant professor at the Hong Kong University of Science and Technology and head of World Mind Lab, working on world models, 3D perception and physical AI. He has published more than 50 papers at venues including SIGGRAPH, CVPR, ICCV and TPAMI.

Another core member, Ji Mazeyu, graduated from the University of California, San Diego, with research spanning 3D perception, dexterous manipulation and whole-body control for humanoid robots. He worked on real-robot validation of a general-purpose humanoid controller and was a founding member of Silicon Valley physical AI company Assured Robot Intelligence, later acquired by Meta.

Simate now has a stack that runs from foundation models through engineering implementation to deployment on real robots, supporting its Physical RSI route.

General-Purpose Physical Fast System Tops RoboDojo

Simate says it is developing physical intelligence models through AutoResearch under a weak-RSI, human-in-the-loop paradigm, and that its general-purpose physical fast-system model has taken first place in the RoboDojo benchmark.

RoboDojo evaluates general robot manipulation across several dimensions, including task generalization, environmental memory, fine manipulation, long-horizon execution and open-ended tasks, giving a fairly broad read on a robot model's overall capability.

In real-robot demonstrations, Simate showed a robot drawing on prior information to complete multi-step tasks such as brewing tea.

Long-horizon tasks require more than reading the current frame. The robot has to remember which steps it has already completed, track progress, and choose its next action based on how the environment has changed. Relying only on the current frame can leave the model short of the information it needs to continue; storing every past observation adds real-time compute load.

Simate's answer was to build 4D physical perception and a memory mechanism into the general-purpose physical fast system. The model interprets its physical surroundings using depth, geometry, motion and contact information, and organizes past information around task progress, continuous state and immediate feedback, seeking a balance between long-horizon stability and real-time responsiveness.

To run larger models on the robot itself, Simate has also explored efficient visual encoding, spatiotemporal modeling and feature compression. The plan is for the general-purpose physical fast system to eventually work alongside a general-purpose slow reasoning system, pushing further into zero-shot and few-shot general manipulation.

AutoResearch Puts AI Inside Robot Development

Simate's other major bet is bringing AI agents into the robot model development process itself.

Researchers have traditionally had to revise models, adjust data, run training and evaluation, then decide on the next optimization step from the results. AutoResearch, developed by Simate, tries to stitch those steps together and hand more of the execution work to agents.

Given a research objective and constraints, AutoResearch can read model code, training configurations and past experiment records, modify model components and settings through SiPAI, call on infrastructure to run training and evaluation, and decide from the results whether to continue, adjust or stop.

The system is more than a bundle of experiment scripts. It manages separate experiment branches, tracks evaluation results across scenarios, and gradually consolidates multiple rounds of experiments into a more stable policy.

With the experiment loop automated, researchers can cut back on repetitive engineering and spend more time on evaluating hypotheses, choosing research directions and analyzing complex results.

The company says AutoResearch is now open for trial use, with researchers at Tsinghua University, MIT and HKUST among those using the platform.

Physical RSI: Robots That Keep Improving Themselves

The RSI in Physical RSI stands for recursive self-improvement. The question it poses is whether an intelligent system can use its existing capabilities to improve itself, and whether those improvements then strengthen its next round of research and learning.

Simate divides RSI into weak, medium and strong tiers based on how complex the research task is.

Weak RSI covers problems with clear boundaries and short feedback cycles, such as revising data processing logic, optimizing a function, or adjusting a computation pipeline under fixed constraints. The system can handle the modification, testing and feedback on its own.

Medium RSI covers human-in-the-loop research that requires multi-stage planning. Here the agent handles context organization, code implementation and moving experiments forward, while researchers screen hypotheses, analyze contradictory results and make the key calls.

Strong RSI takes on more open-ended problems, such as improving a model's generalization to new tasks. The system has to understand the objective itself, identify questions worth investigating, and hold a research direction over longer periods while correcting its technical approach. For now, that tier still needs senior researchers in the lead.

Simate argues all three forms of RSI can coexist within a single development effort. Its long-term goal is to steadily reduce human involvement in repetitive research work, letting models, tools and methods improve continuously and strengthen each subsequent round of research.

Moving from weak to strong RSI hinges on whether agents can go beyond running predefined experiments to independently spotting problems, forming useful hypotheses and judging research direction. That is the line separating automated research from research with a higher degree of autonomy.

From Simulated Feedback to Real-Robot Validation

To sustain Physical RSI iteration, Simate has built out a stack covering model frameworks, data processing, automated research and real-world deployment.

SiPAI is a pluggable model framework whose module boundaries, interfaces and validation flows are designed so agents can understand model structure, locate the components that need changing, and run standardized training and evaluation once changes are made.

SiPAI supports architectures including world models, world action models, vision-language-action models and vision-language models. If the goal is better memory, for instance, the agent can work on the memory module; if the issue is data mix, it can adjust the training configuration.

Ultimately, though, gains in robot model capability still have to be proven in real operation.

Simate uses world models and simulation environments for fast experimental feedback, then tests model performance on real robots, feeding failure cases from real operation back into the research process.

In parallel, AutoResearch validates research hypotheses through model and data experiments, uses evaluation results to identify capability bottlenecks, and drives the next round of data collection and model training, closing the loop from model modification to experimental evaluation to real-robot validation to feedback iteration.

AI-Driven Robot Development Enters a New Phase

As data, models and computing power keep expanding, the robot development process itself is starting to scale and automate.

Simate's entry point is AutoResearch: AI agents take part in model development, experimentation and evaluation, and feedback from real robot operation becomes the basis for the next round of research. The company has opened recruitment for a closed beta of the AutoResearch platform.

If the Physical RSI route matures, robot development could shift from repeated human trial and error toward AI-assisted model modification, experiment design, performance evaluation and policy iteration, lowering development and deployment costs while making robots better at adapting to complex tasks.

Whether AI can genuinely propose useful hypotheses, identify the key problems and keep robot capabilities evolving on its own, however, still needs more public benchmarking and validation in real-world settings.