Simate-beta Debuts as the Top-Performing Physical AI Model on RoboDojo

AI studies AI, and this time it has an answer from the physical world.

A company founded just three months ago has unveiled its first general-purpose physical agent system, Simate-beta: demonstrating memory, long-horizon tasks, fine-grained manipulation, and task adaptability—and, according to the company, has topped RoboDojo.

The leaderboard, of course, does not represent the full capabilities of a model, but it provides a window into the pace and strength of development.

According to the team, the base model submitted for evaluation was not specifically optimized for this ranking.

After only three months, achieving such results with the first version of the model is noteworthy.

The team is called Simate (Silicon Mate).

The team also has impressive credentials. According to the company, the core team successfully advanced a single-stage end-to-end autonomous driving model to a level comparable to Tesla FSD and achieved mass production deployment.

Globally, very few teams have crossed this threshold; even fewer continue advancing general physical AI with this accumulated expertise.

However, the most interesting aspect this time is their development approach, established from day one:

AI for Physical AI: Enable AI to participate in the research and iteration of physical intelligence itself.

The company has organized its models, data, computing power, and experimental workflows around this approach. Simate-beta is the first practical outcome of this AI-native R&D system.

Moreover, research tools have already been adopted by external researchers.

According to the company, researchers from MIT, Caltech, Tsinghua University, Peking University, and other institutions have participated in the beta testing of the AutoResearch platform. Its Sinfra infrastructure supports training, simulation, and inference; available resources are confirmed upon application.

Meanwhile, Simate has successfully completed multiple funding rounds amounting to hundreds of millions of RMB.

Over three months, the model, research platform, external trials, and fundraising are progressing in parallel. This pace certainly makes one eager to see what’s next.

The team has also provided a preview: they have developed a technical approach for zero-shot generalization on complex tasks and plan to release phased results by the end of this year.

In their view, GPT-6 beginning to engage in physical world operations is an expected advancement, and breakthroughs in Physical AI may come faster than outsiders imagine. They aim to further move beyond reliance on task-specific adaptation and post-training, propelling Physical AI toward its own "GPT-3 moment."

Three research initiatives, including model and automation research, will be gradually published through papers and technical reports, with related results planned for open-sourcing in phases.

The speed of this development, and whether it can be sustained, becomes the most important thing to watch for Simate next.

So, what exactly did the first model accomplish? And how did AI participate in the research?

Simate ultimately aims to achieve zero-shot and few-shot general physical AI operations.

In a completely new and unfamiliar environment, the system should not collect new data, retrain the model, and re-tune engineering parameters for every new task added.

This is precisely the most popular and challenging field in physical AI today—where numerous players compete, each showcasing their unique strengths.

The solution provided by Simate at this stage is called the Universal Physical Fast System.

That is System One.

Let’s start with a human example:

When faced with complex instructions like “Help me tidy up the table,” we need “System 2” to understand the intent and break down the steps;

But when your hand is already reaching for the cup, deciding whether to adjust a few millimeters left or right—these millisecond-level reactions can’t possibly rely on deliberate, step-by-step thinking.

The same applies to physical AI.

A general-purpose "slow system" handles high-level planning and complex reasoning, while a sufficiently powerful "fast system" perceives the ever-changing physical world in real time and outputs actions at high speed.

Simate chose the latter—

Increase the parameter scale of the fast system to explore the emergence of zero-shot generalization capabilities and push the limits of size and performance for edge-deployable models.

A general physical fast system.

However, big and fast is a difficult trade-off to manage.

Simate chose to start with two things:

1. 4D Physical Sensing

To understand the physical world, one must simultaneously capture the dynamic changes in spatial structure and time dimension, ensuring that physical representations truly serve action.

2. Hierarchical Temporal Memory

Long-term tasks require remembering past events, continuous operations need to track state, and immediate responses must not be slowed down by large volumes of historical data.

4D physical perception and hierarchical temporal memory align together with how the human fast system processes the continuous world.

This approach also builds on the team’s expertise in large-scale vision models, temporal modeling, real-time inference, and production deployment. The specific architecture and model scale will be disclosed in subsequent technical reports.

The desired end state is for the model to clearly understand what is happening right now, remember what just occurred, and make timely decisions about the next step.

Collaborate with general reasoning capabilities represented by cutting-edge models like GPT-6 to advance zero-shot execution of complex tasks.

Based on the current Simate-beta physical demonstration provided by the company, testing primarily focuses on task adaptation driven by demonstration, memory, complex long-term execution, and fine-grained operations.

According to the RoboDojo ranking provided by the company and updated on September 23, 2026, simate-beta ranks first with an average Score of 33.95, corresponding to an SR of 27.96%. This result is based on the ranking evaluation; actual device performance is displayed separately.

Behind this model answer lies another system equally worthy of exploration.

In recent years, people have become increasingly accustomed to AI writing code.

But in AI research, writing code is only a small part.

Often, iteration efficiency is the key to victory. Whoever can test more research hypotheses within a given time frame will emerge first.

Here's the simplest example.

Long-term tasks for the bot are always stuck at a certain stage, and the reasons are typically one of the following: data distribution distortion, model architecture bottlenecks, trajectory imbalance, or ineffective training strategies.

Behind every direction are dozens or even hundreds of experiments.

If we relied entirely on researchers to manually adjust configurations, launch tasks, monitor training, run evaluations, and then compile the results, the opportunity would be long gone...

So Simate created AutoResearch to accelerate this process—

Human researcher: Formulate hypotheses, set objectives, define constraints.

AutoResearch Engine: Automatically take over to decompose, execute, and automate the entire process of feedback for the remaining experiments.

The final results have been revealed—they secured first place at RoboDojo using this method.

More importantly, speed.

Simate took only three months from its founding to reaching the top of RoboDojo.

That's just ridiculous...

So the question arises:

How should AI "research AI"?

Clearly, sending in just one agent is far from sufficient.

True robotics development penetrates the underlying layers of model code, data pipelines, training clusters, evaluation environments, real-machine feedback, and experimental records spanning months or even years.

Ultimately, this is still a matter of "context." The agent needs a super interface that can coordinate everything.

To this end, Simate has built a three-tier architecture: SiPAI + AutoResearch + AI-native Infra.

SiPAI: Making model architectures "AI-researchable"

This is a plug-in model framework natively designed for automated research.

In simple terms, it means building the robot model into a structure that AI can easily understand—module boundaries, system interfaces, configuration items, and validation processes are all defined with extreme clarity.

Only in this way, after receiving the research topic, can the Agent precisely determine: which module to modify, what areas will be affected by the change, and how to verify the outcome ultimately.

Simate has now completed the construction of this infrastructure and successfully integrated with major architectures including World Models, World Action Models, VLA, and VLM.

Both human researchers and agents can freely combine components, adjust local implementations, and use the unified training and evaluation pipeline.

AutoResearch: Provides continuously updated "research context"

If SiPAI provides a static manual, then AutoResearch addresses the dynamic information gap:

How far has the research progressed? What pitfalls have been encountered before? What is the latest evidence?

It integrates external cutting-edge papers, internal team discussion materials, model code, data ratios, historical experiment logs, and the latest evaluation feedback into a single dynamic research context.

This Context will be updated in real time, whether a new paper has been published, a new version of internal code has been submitted, or the latest experimental data has overturned previous assumptions.

This way, each research suggestion proposed by the Agent can be precisely anchored to a specific code branch and experimental version:

The most fascinating aspect is the deep integration between "human experience" and "agent experiments":

Human researchers can随时 inject constraints and optimization directions; once an experiment validates an effective component, it immediately becomes a new starting point for subsequent experiments; methodologies developed in one branch can be seamlessly reused by the next Agent.

The research findings have truly become the means of production for the next round of research.

AI-native Infrastructure: Keep the 24/7 R&D flywheel spinning

Of course, an agent shouldn't just write code—the experiments must run efficiently.

Simate integrates the entire workflow of training, inference, and evaluation into its proprietary infrastructure, enabling parallel execution of dozens of independent research pipelines through optimized task orchestration and resource scheduling.

To prevent computational waste, Simate employs a high-frequency filtering mechanism: all routes are first subjected to an initial rapid filter through the "world model + simulation environment," eliminating the vast majority of ineffective directions;

Only truly promising solutions will proceed to real-world testing.

New issues have surfaced on real devices, prompting a return to the research context, thus closing the loop:

The final result is that humans only need to propose an exploration direction, and the underlying system automatically runs dozens of experiments in parallel at high speed.

This is also the key factor that enabled Simate to rise to the top within three months.

But most surprisingly, they chose to directly open-source this highly core internal productivity tool!

The webpage looks like this; interested users can get it at the end of the post.

Researchers from top universities such as Tsinghua University, MIT, and HKUST have already joined the beta testing.

Founder Zhang Ying, former core technology lead at a top-tier autonomous driving company.

Extensive hands-on engineering experience, having contributed to the development of three generations of intelligent driving solutions—with and without images, and end-to-end—and having extensive experience in mass production.

According to reports, the autonomous driving system it has developed is the closest in China to Tesla's FSD.

In addition, there are two key figures on the team.

1. Occupancy rights.

Assistant Professor at the Hong Kong University of Science and Technology and Head of World Mind Lab, with long-term research focus on world models and physical AI.

2. Ji Ma Zeyu.

A post-2000s young scientist whose research spans 3D perception, dexterous manipulation, and whole-body control of humanoid robots.

Before joining Simate, he was a founding member of Assured Robot Intelligence in Silicon Valley.

ARI was later acquired by Meta.

The capital markets have also voted with real money.

Just three months after its establishment, Simate has completed multiple funding rounds, each amounting to hundreds of millions of RMB.

And all of this is a bet on the same thing—

Physical RSI.

From its very first day, Simate targeted this direction and defined three different types of RSI.

1. Weak SI: Clearly defined boundaries, rapid closure

Human-machine collaborative mode: Researchers set goals and constraints; no further step-by-step involvement is required during normal operation, and exceptions can be escalated for handling.

This ascent of RoboDojo marks the first phase of engineering validation.

2. CSI: Clear direction, but unknown outcome

How can model memory be improved? How should the data ratio for a specific type of task be configured?

The direction has been established, but determining which specific approach works best requires multiple rounds of testing.

At this stage, the agent is responsible for organizing context, implementing solutions, running experiments, and comparing results; the researcher is responsible for filtering hypotheses, interpreting conflicting evidence, and making critical trade-offs.

This is also the type with the most practical value at this stage.

3. Strong SI: The research itself must be discovered

The most difficult—and most fundamental to the essence of intelligence—type of—

How can I improve a model's generalization ability on entirely new tasks?

There is no longer a clear path to solving these types of problems. The system must understand the goal on its own, identify truly worthwhile research questions, and maintain direction over a long period while continuously adjusting its course.

This begins to touch on what is known as "research taste."

At this stage, such work still requires senior researchers to lead, with agents primarily providing support in literature review, evidence analysis, experimentation, and pathway exploration.

It is worth noting that the weak, medium, and strong SI states mentioned by Simate are not hierarchical; all three can coexist within the same development system.

More importantly, it's the first principle of "evolution" itself.

Each self-improvement of the system makes the next improvement easier to achieve.

When data, models, and real-world deployment form a self-sustaining loop, the upper limit of robots is no longer constrained solely by human limitations.

This is Physical RSI, currently the most popular candidate pathway for physical AGI.

Of course, what Simate is currently showing is still in its early stage.

Human-in-the-loop still exists, and researchers still hold the key directional judgments throughout the entire cycle.

But this also shows that automated research doesn't need to wait until the day of "complete human absence" to begin generating value.

Before truly advanced autonomous research capabilities emerge, gradually delegating high-cost, processable tasks such as experiment design, training, evaluation, and feedback to agents can already significantly improve research throughput.

RoboDojo's transcript also serves as validation of the "human-machine collaboration" approach in the field of physical AI—

RSI has really begun.

By the way, there's one more thing.

The AutoResearch research platform has launched its beta testing and is currently actively recruiting participants!

Thousand kcal of computing power, come and play with us~

Official website:

https://mate-robot.cn/

AutoResearch Research Platform:

https://mate-robot.cn/research/sinfra/

This article is from the WeChat public account "Quantum Bit" (ID: QbitAI), authored by Jay.