In the first half of 2026, the embodied AI track entered a subtle stage.
On one hand, there is a carnival of capital and media—almost every week, new robot hardware companies are unveiled with financing, showcasing bipedal walking, dexterous grasping, folding clothes, and pouring water, with videos becoming more and more flashy. On the other hand, those who are truly grinding in factories and laboratories know that there is a huge gap between these demos and being 'usable'.
Where is the problem? The controversy is not about whether humanoid robots are necessary; it boils down to three things: where the data comes from, how the brain thinks, and how the hands and feet move.
The technical route for the 'brain' is far from converging. Papers on so-called 'World Models' are everywhere, but those that can actually run on robots and interact with the physical world are extremely rare.
A few days ago, Fei-Fei Li announced that World Labs acquired SceniX, entering the field of embodied AI simulation. She posted that the world model is not just a generative model, but is more reflected in its interaction with the physical world.
Yes, the world model in the field of embodied AI is the 'brain' model of embodied AI, not those cool generative AI models, which are just AI models in the digital world.
At the beginning of this year, a dark horse company emerged in the field of embodied AI: Synkrotron. In just half a year, it has enabled embodied AI robots to have actual landing scenarios, run through the business closed loop, and even sent robots to work in scientific research laboratories with extremely high requirements for complexity and precision.
Frankly speaking, this speed is somewhat abnormal. It usually takes three to five years for a company to go from scratch to scenario implementation. What is different behind running through the closed loop in just half a year?
(I) Understanding Science, Excelling in Engineering, Mastering Management, and Strong Execution
Let's start with the team. This might be the place where Synkrotron is most easily labeled by the outside world—incubated by Academician Andrew Yao, the 'Turing Award' winner.
No need to introduce who Andrew Yao is: Turing Award winner, Dean of the Institute for AI (Artificial Intelligence) at Tsinghua University, Academician of the Chinese Academy of Sciences, and the founder of the 'Yao Class'.
The team configuration of this company, in their own words, is 'understanding science, excelling in engineering, mastering management, and strong execution.' This is not just a slogan, but a fact you will discover after laying out the resumes of several core members.
The composition of Professor Zijiang Yang's team, the founder of Synkrotron, presents a rare structure of 'academic depth × engineering density.' Dr. Zijiang Yang is the founder of Synkrotron, having published over a hundred papers and holding more than 20 patents in China and the US. He is also a professor and doctoral supervisor at the School of Computer Science and Technology, University of Science and Technology of China (USTC), Director of the Embodied Cognitive Science and Spatial Interaction Engineering Technology Center at the USTC Institute of Advanced Technology, Vice Chairman of the IEEE Autonomous Driving International Standard Working Group, and Chairman of the IEEE AI Agent Standard Working Group (under establishment).
President Yurong Zhou has 25 years of experience as an executive in technology enterprises, having served as the President of Laiye Technology, General Manager of VMware's Greater China Commercial Clients, and General Manager of Symantec's Northern and Greater China. He is a veteran deeply engaged in the enterprise-level market. CTO Dr. Yuxi Pan formerly served as the Technical Lead and Senior Data Scientist for Uber's Data Science Platform Group, accumulating rich industrial-grade system practical experience at first-line technology companies such as Cisco's Security Data Group and Singularity Energy. Chief Scientist Dr. Xiaodong Zhang formerly worked as an engineer at Huawei and has published more than 20 high-level papers in the fields of software engineering and AI.
With such a team, it is hard to summarize it with a single label. There are those who do theory, engineering, products, and business. In the track of embodied AI, which requires simultaneously crossing the three hurdles of scientific cognition, engineering implementation, and commercial landing, the scarcity of a comprehensive team is far more valuable than single-point excellence.
Many people think that starting a business means having a core technology to conquer the world. But technology is just a ticket to enter. What truly determines whether a robot company can land scenarios in half a year is whether there is someone in the team who knows 'how to move a technology from the laboratory to the factory, and then from the factory to the customer.' This kind of ability cannot be learned by reading papers.
(II) 'Dimension Upgrading/Downgrading' from Autonomous Driving to Embodied AI
The autonomous driving background of the core team is actually the most underestimated dimension of Synkrotron.
Why is the experience in autonomous driving so important for making robots?
Autonomous driving is essentially 'robots + open world,' and it is also the first landing scenario for robots interacting with the physical world. An autonomous driving car needs to perceive, make decisions, and control in real-time in a complex physical environment. This is highly isomorphic to the underlying logic of an embodied AI robot pouring test tubes in a laboratory or moving boxes in a warehouse.
The Synkrotron team has worked on everything from sensor simulation, multi-agent behavior modeling, and vehicle dynamics to end-to-end models. This means they have already gone through the pitfalls of 'how to use synthetic data to train a system that runs in the real world.'
The 'data famine' and the 'Sim2Real Gap' currently faced by embodied AI were already experienced by the autonomous driving industry between 2018 and 2022. The solution at that time was: use high-fidelity physics simulation engines to generate massive synthetic data, covering various long-tail scenarios, and then gradually migrate to real deployment.
Synkrotron's 1.0 business model was precisely to provide physically accurate synthetic data for autonomous driving—which means they already had mature engineering accumulation in the matter of simulation engines. Now, migrating the same capabilities to robots is not starting from scratch, but 'upgrading the dimension' of the methodology already verified in the autonomous driving track to embodied AI.
However, the 'dimension upgrading' is only reflected in the complexity of the robot; in some other aspects, it is even considered 'dimension downgrading.'
The requirements for real-time performance, safety, and data quality in autonomous driving are higher than those of robots in many dimensions. A system that needs to ensure millisecond-level response at a speed of 120 km/h, when migrated to a robotic arm slowly pouring reagents in a laboratory, is downgraded by more than one order of magnitude.
This is why Synkrotron can run through the closed loop in half a year. Not because they run fast, but because they are standing on a path that has already been successfully walked through.
(III) World Model Data Generation Under Physical Constraints
Having talked about the 'people,' let's talk about the 'food.'
What is the scarcest resource in embodied AI? It is not computing power, nor algorithms, but data.
The data mentioned here is not ordinary images or text, but the operational data of robots interacting with the environment in the physical world—how the hands grasp, how the force is controlled, and how the pose of objects changes in space. The cost of collecting such data is extremely high: it requires real robotic arms, professional operators, and repeated trial and error. A high-quality operational trajectory might cost hundreds of dollars. Moreover, many long-tail scenarios—such as grasping transparent objects under extreme lighting—are impossible to collect even if you want to.
The solution provided by Synkrotron is a 'World Model Data Generation Platform Under Physical Constraints.' In plain language, it means: using a simulation engine combined with a world model to jointly drive the batch generation of photo-realistic training data.
Key technical indicators of this solution include:
The visual realism FID (Fréchet Inception Distance, an indicator measuring the distribution gap between generated images and real images) of the generated data is close to the real data distribution, meaning the domain gap is significantly compressed.
It supports combinatorial generation of rare object interactions, extreme lighting, and complex occlusion scenarios—which is the so-called 'infinite expansion of long-tail scenarios.'
The physical simulation engine performs constraint verification to ensure that the generated data is physically consistent.
Coupled with their first-person data collection (Ego-centric) equipment and UMI-based body-free crowdsourcing collection solutions, a data pipeline with a 10x cost reduction is built—no robots, no professional operators, and not even a fixed laboratory is needed. Anyone in the world can collect data using head-mounted devices.
Combining virtual and real, synthetic data + real robot data + Ego data, three data sources flow into a unified data lake, forming a closed loop. This is actually the core logic they repeatedly emphasize in their BP: breaking the bottleneck of data costs in the real world.
(IV) Dora: A Native Foundation Designed Specifically for Embodied AI
When it comes to robot operating systems, the whole industry cannot bypass one name: ROS (Robot Operating System, robot operating system and middleware).
ROS was born in 2007, an open-source framework developed by Stanford University and Willow Garage. For over a decade, almost everyone in the world making robots has been using it. But the problem is that ROS was never designed with the needs of the large model era in mind—a VLM (Vision-Language Model) running on a robot often has tens of gigabytes of parameters, and high-frequency multi-modal sensor data streams pour in. The data transmission architecture of ROS simply cannot handle it.
The answer given by Synkrotron is Dora.
Dora is not just another 'improved version of ROS.' It is a new generation of robot-native operating system launched in 2022 by Synkrotron jointly with Huawei Futurewei, specifically designed for world models and embodied AI.
Thomas Wolf, co-founder and chief scientist of Hugging Face, commented after seeing Dora: 'A super impressive replacement of ROS.'
The weight of this evaluation is quite heavy in the AI open-source circle.
There are two core optimizations at the bottom layer of Dora, which are worth breaking down.
First, direct pass-through of large model tensors. Traditional ROS2 requires serialization/deserialization when transmitting data—simply put, data is packed into byte streams at the sending end and unpacked at the receiving end, which incurs considerable overhead. In the era of large models, the action control tensors output by WAM models often have millions of parameters. If they go through the traditional serialization path, the latency is simply unacceptable.
Dora's approach is to break memory isolation—the memory data format of the tensors output by the model is completely consistent with that of the hardware control nodes, enabling direct memory mapping and zero-copy transmission. Measured data: the transmission latency of Dora is 0.9 microseconds, while ROS2 is 15.2 microseconds, making data transmission efficiency up to 17 times faster.
What is the concept of 0.9 microseconds? It is one-seventeenth of ROS2. In a system that requires high-frequency real-time inflow of sensor data, real-time inference of large models, and real-time distribution of motion control, this latency gap could be the difference between 'being able to run' and 'not being able to run.'
Second, deterministic scheduling of heterogeneous computing power. Large models run on GPUs, sensor collection runs on CPUs, and motion control also runs on CPUs—the three are competing for computing power. Dora provides kernel-level deep customization and CPU core isolation (CPU Isolation) to ensure that the computing power for critical tasks is not squeezed by large model inference.
Putting these two optimizations together solves the most core contradiction of the embodied AI operating system: large models need to consume a lot of computing power, while real-time control requires deterministic low latency. Traditional solutions can only compromise in the face of this contradiction—either reduce the model inference frequency or sacrifice control accuracy. Dora eliminates this compromise from the underlying architecture.
Hardware-software co-optimization is the true meaning of the word 'foundation.'
(V) WAM: The World Action Model That Lets Robots 'Understand' the Physical World
With data and the operating system, the last piece of the puzzle is the 'brain'—which is Synkrotron's WAM (World Action Model).
Currently, there are three mainstream routes for making the robot 'brain': the end-to-end scheme driven purely by large models, the modular perception-planning-control architecture, and the world model route. The problems with the first two are obvious—the pure large model scheme lacks safety and determinism in the open physical world, and the modular scheme has obvious bottlenecks in cross-scenario generalization.
Synkrotron takes the world model route but makes a key design: a unified world action model.
In traditional schemes, 'understanding the world' and 'outputting actions' are two isolated modules—the perception module looks at the scene and outputs a state, and the decision module takes this state to plan actions. Information transfer between the two modules inevitably has losses.
WAM integrates these two things into one model: it performs both action generation and state prediction simultaneously. Inputting the current state and task goal, the model not only outputs 'what action to take next' but also synchronously predicts 'what the world will look like after doing this action'—forming a closed loop of 'perception—decision—prediction.'
The advantage of this design is that the robot is no longer 'blindly executing.' It has already rehearsed the consequences of the action before making every move. If the prediction results show 'hit the wall' or 'dropped the cup,' it can adjust in advance. This is the so-called 'safe interaction in dynamic open environments.'
(VI) Entering the USTC Laboratory: First Landing in Scientific Research Scenarios
Having said so much about technology, it ultimately comes down to implementation.
The first stop chosen by Synkrotron is not a factory or a warehouse, but the National Key Laboratory of Precision Intelligent Chemistry at the University of Science and Technology of China (USTC), a scenario with extremely high requirements for operational complexity and accuracy.
This choice is quite interesting.
Why a scientific research laboratory? Because AI for Science is becoming a global consensus—in 2025, the US launched the 'Genesis Project' to build a runnable 'Scientific AI' closed-loop system within 9 months; the UK also listed AI4S tools and autonomous laboratories as key capability construction directions in its AI science strategy.
China is also promoting this. In April 2025, the top leadership explicitly proposed 'leading the reform of the scientific research paradigm with AI, accelerating breakthroughs in scientific and technological innovation in various fields, and building a collaborative innovation system of industry-university-research-application led by enterprises.'
Scientific research experiments are a highly structured scenario with standardized operational processes, where every step has clear physical and chemical rule constraints. For a world model-driven robot, this is almost a perfect 'training ground': task boundaries are clear, safety constraints are explicit, and the results of every operation can be judged by quantifiable indicators.
The cooperation between Synkrotron and USTC, implemented at the specific level, is as follows:
The robot adopts Synkrotron's self-developed Dora operating system, replacing the laboratory's original robot system based on foreign ROS—which in itself is a move for 'domestic substitution.' Then, it integrates the AI intelligent vision algorithm of the 'Lingjing' embodied AI agent to visually monitor every step of the robot's operation, ensuring reliability. More importantly, both parties are jointly building a 'Scientific Experiment World Model'—integrating physics and chemistry knowledge, so that the robot not only mechanically executes preset actions but can understand the principles of the experiment, possessing the ability to reason and autonomously execute complex experimental tasks.
The laboratory requires the robot to have three capabilities: motion and operation capabilities (autonomous movement, long-horizon task planning, loading and unloading, flexible instrument operation), experimental exploration capabilities (imitating and learning experimental skills, operating precisely like an experimentalist), and spatial interaction capabilities (collaborating and communicating with scientific researchers, understanding scientific research needs and translating them into instructions).
These three capabilities exactly correspond to the outputs of WAM, Dora, and the hybrid data platform.
In other words, Synkrotron's three core products—hybrid data, Dora operating system, and WAM brain robot—have formed a complete verification closed loop in the scientific research laboratory scenario. This is not a demo, but a truly running system.
In addition to USTC, Synkrotron is also building a joint innovation center with Xinghaitu and Linghou Robot, conducting technical joint debugging and standard co-construction, with the goal of promoting domestic embodied AI solutions.
(VII) In Conclusion
Returning to the question at the beginning: why can this company land robot applications in half a year?
The answer is actually not complicated.
It is not because they run faster than others, but because they did not start from scratch at the very beginning. The academic foundation incubated by Academician Andrew Yao, the simulation data and systems engineering methodology accumulated over more than ten years in the autonomous driving industry, and a comprehensive team that 'understands science, excels in engineering, masters management, and executes strongly'—the superposition of these three things determines that Synkrotron was already one step ahead at the starting line.
Then there is the choice of technical route. The Dora operating system solves the problem of 'how the hands and feet move'—17 times the transmission efficiency of ROS2, and kernel-level deterministic scheduling of computing power. World model data generation under physical constraints solves the problem of 'where the data comes from'—photo-realistic, physically consistent, and infinitely expandable for long-tail scenarios. The WAM world action model solves the problem of 'how the brain thinks'—joint prediction of action and state, VLM coach feedback, and continuous evolution.
Putting the three pieces of the puzzle together forms a complete embodied AI technology stack.
Finally, there is the landing strategy. Choosing to cut in from scientific research scenarios instead of plunging headfirst into the red sea of industry and commerce is a smart choice. Scientific research laboratories have a high degree of standardization, clear safety constraints, and quantifiable results, making them the most ideal transition zone for robots to go from 'being able to run demos' to 'truly working.' After running the system stably in the laboratory, expanding to industrial, medical, and commercial scenarios will naturally follow a smooth path.
According to their plan, by the end of 2026, they will achieve the development of the 'Human-Robot Collaboration World Action Model,' completing the sales and scenario landing of 100 most powerful brain hardware bodies; by the end of 2027, they expect to sell 1,000 units; and the target for 2028 is 10,000 units. Scientific collaboration robot products will simultaneously be piloted in universities and the medical field.
Whether these numerical results will meet expectations or exceed them, no one can tell for sure now. But at least at the current node, Synkrotron has shown the industry a possibility: the landing of embodied AI does not necessarily have to start with making a cool hardware body. First, solidify the three things—data, operating system, and brain—and landing will be a natural consequence.
Bu Rixin, founder of Chuangdao Hard Tech, Vice President of the IC (Integrated Circuit) Professional Committee of Beijing University of Posts and Telecommunications, expert at Zheshang Securities Research Institute, and concurrently serving on the investment committees of multiple investment institutions. He holds a bachelor's degree in Electronic Engineering and a master's degree in Computer Science. He possesses securities and fund qualifications and has over 15 years of experience in IT R&D, consulting, investment, and financing. His investment focus areas include semiconductors, intelligent manufacturing, and NEV (New Energy Vehicle), etc. The 'Hard Tech New Force' platform covers over 50,000 investors and thousands of technology start-ups, and has guided dozens of technology enterprises in completing equity financing.