Cewu Lu is one of the early pioneers of embodied intelligence in China. His academic path began in 2002 with communications and information systems, followed by further study at the Chinese Academy of Sciences and the Chinese University of Hong Kong. These experiences laid the foundation for his later work in embodied intelligence. Although his career has crossed several disciplines, he believes the broad perspective and rigorous thinking cultivated by the information sciences have been essential to his research. This interview retains its January 2025 context.
In 2013, he began working on 3D vision, exploring how to connect two-dimensional images with the three-dimensional physical world and laying groundwork for advances in robot intelligence. From 2015 to 2016, he worked at the Stanford AI Lab on early explorations of embodied intelligence. The field attracted relatively little attention then, but he and his colleagues remained enthusiastic and willing to experiment.
After returning to China, Lu became a professor at Shanghai Jiao Tong University and co-founded Flexiv with Shiquan Wang. That entrepreneurial experience helped prepare him to establish Noematrix. He believes moving from research to commercialization requires a wide range of capabilities, and hands-on experience gave him a deeper understanding of the entire journey from technology to practical value.
Today, Lu is a leading figure in China's embodied-intelligence community and has made significant progress in integrating AI with robotics. His goal is to help robots become part of human life and serve as true partners.
This episode of ReSpark features Noematrix founder Cewu Lu and investor Tim. The following is the written account of their conversation.
Original article outline
Interview highlights
Part I: Personal experience
1. Academic background
2. The Flexiv experience
Part II: Noematrix
3. Noematrix's approach to data
4. Collecting robot data
5. Open robot datasets
6. Robot force control
7. Noematrix's embodied foundation models
8. Progress on Noematrix Brain
9. Solving long-horizon tasks
10. Adapting to different hardware
11. Combining vision and touch
12. Commercial deployment
ReSpark: You started your undergraduate studies in 2002 and, as I describe you in this question, were born after 1985, yet you are already one of China's most experienced professors in embodied intelligence. You initially studied communications and information systems, then pursued a master's in electromagnetic fields and microwave technology at the Chinese Academy of Sciences. Was your PhD in computer science and engineering at the Chinese University of Hong Kong the point when your studies first became directly connected to your present work? Editorial note: the birth-year remark is retained from the question. Lu's university profile confirms the education dates but does not give his birth year.
Cewu Lu: Fundamentally, embodied intelligence is still a form of artificial intelligence, so my early AI training was enormously helpful. Taking a broader view, AI itself emerged from the development of the information sciences. Studying those disciplines emphasizes an overall perspective, careful scientific reasoning, and complete training from theory through engineering. On the surface, some fields or technologies may seem far apart. Looking back, however, you often discover that those early foundations have quietly supported your later work.
ReSpark: You went to the Hong Kong University of Science and Technology in 2013 for postdoctoral research. What were you studying?
Cewu Lu: My research was already gradually becoming connected to embodied intelligence. I mainly worked on 3D vision because I wanted to move from studying two-dimensional images toward the real physical world, and 3D vision was the most direct entry point. If you look at my papers from that period, many concern 3D vision and understanding the three-dimensional world. Some were formally published after I left HKUST, but the chronology shows that I was already exploring how to move from the world of images into the physical world.
ReSpark: From 2015 to 2016, you did postdoctoral research at the Stanford AI Lab, working with Leonidas J. Guibas and Fei-Fei Li. Are there interesting experiences or impressions from that time that you could share?
Cewu Lu: I do not think there were distinct camps such as a 'simulation school' then. Everyone recognized embodied intelligence as a direction for the future and was eager to get started, but thinking and experimentation were still at a very early stage. If a method could produce results, we were not committed to any one school of thought.
The atmosphere was fascinating. A group of young people who believed this was the future came together to discuss wild ideas and run experiments. Many experiments failed, of course, but that did not diminish our enthusiasm. The wider world was paying attention to autonomous driving and intelligent healthcare, yet many people did not understand the robotics or embodied-intelligence work we were doing.
I remember that publishing a paper connected to this kind of robotics or AI work could be difficult, and outsiders were unfamiliar with the new ideas. Within the group, though, we enjoyed doing frontier research in a small community. It was a little like early philosophical inquiry: we did not merely imagine possibilities; we also tried to put them into practice.
Shiquan Wang, who was in the mechanical engineering department, and I tried many ideas together. People around us did not necessarily understand what we were doing, but we enjoyed ourselves and developed bold ideas. Some of those ideas have gradually become reality, and I believe more will follow.
ReSpark: After leaving Stanford, you taught at Shanghai Jiao Tong University while building Flexiv with Shiquan Wang. How did that experience help you when you went on to establish Noematrix?
Cewu Lu: It helped enormously. Turning research into an industrial product demands a very broad set of capabilities that books alone cannot teach. Professors may be able to achieve the initial research breakthrough, but turning it into a solid, useful product requires strong product thinking. Creating genuinely new value then requires commercial judgment and the abilities of both an entrepreneur and a technology innovator. Team management and organizational development become another major challenge further along the way.
I had read books about these things, but only when I entered the arena myself did I realize how many connections had to be made between the initial research result and ultimately creating value for users. Returning to those books after hands-on experience gave me a deeper, more complete understanding of their lessons.
In building Noematrix, I have worked through research, product development, management and commercial deployment. Every high-tech or AI-driven company has its own story, but similar central challenges and patterns emerge at different stages. I now feel as though part of the map has been illuminated. Challenges remain, but I roughly know where the next 'monster' is and have a clearer idea of how to face it.
Over the past eight years at Flexiv, we have had successes as well as twists, setbacks and frustrations. Acquiring these valuable lessons has required considerable time and expense.
ReSpark: How did you choose the Chinese name Qiongche for Noematrix?
Cewu Lu: It is interesting because it reflects an aspiration shared by our founding team. There are two layers of meaning. The first character, qiong, evokes the sky: the frontier of science and technology has no end, and exploration should continue indefinitely. The second, che, means going all the way through—being truly fundamental, not merely deep, but also grounded in practical implementation. You could describe it as reaching for the heavens while standing firmly on the earth. Reaching upward means extending theory and technology at the frontier; standing on the earth means realizing them in actual applications. Together, they express continual exploration of an unlimited frontier and creating value through deep, practical implementation.
ReSpark: Some people assume Noematrix mainly works with simulation data. Would you like to clarify that? What kinds of data do you use, and when is each type most effective?
Cewu Lu: We should begin by asking what information embodied intelligence needs, rather than deciding in advance which type of data we want. We allocate data according to its informational contribution while considering acquisition cost and whether that cost remains affordable at scale.
We generally place data along two axes: the information it provides and the cost of acquiring and processing it. Finding the right balance produces the best overall result, rather than relying on one source. Our system collects many kinds of data and systematically analyzes their complementary relationships: what useful information each provides, and what uncertainty or noise it might introduce.
The goal is the best overall system performance, rather than declaring a source important or unimportant by intuition. What matters is the result. We certainly use simulation data, for example, but the key is understanding what it contributes, its limitations, and how to combine it with other sources to balance information and cost.
What does simulation data contribute? It has advantages and disadvantages. It helps us build digital assets: creating and manipulating objects in a digital world and simulating their transition from state A to state B. We recently proposed a 'data genes' methodology for constructing and managing these assets. The limitation is that pure simulation often cannot accurately reproduce intensive, sustained physical contact. If we rely on it alone, important information from the real world remains missing.
We therefore also need data collected from real robots or operating hardware. But real-robot data has its own scaling problem. Collection requires continuous human operation, making it expensive and inefficient. Wear, damage or disruption during operation can further limit large-scale collection.
To address this, we proposed a wearable exoskeleton that lets people collect data while doing their everyday work. This can substantially reduce cost and make scaling more feasible. The recorded information can be mapped to the robot, reducing discrepancies between the environment, the person and the machine.
We always return to the broader questions: what information does embodied intelligence need; what does each source contribute; what errors and noise does it introduce; and how can the sources compensate for one another? In our system, simulation provides rules for how the world evolves digitally, real-robot data fills in genuine physical information, and video or human demonstrations provide common-sense knowledge. Like the stages of a rocket, these sources work together: information from earlier stages helps reduce noise in later ones, ultimately producing a stable, complete data system.
We continually focus on scaling both collection and use. The whole system needs to expand sustainably in terms of cost and efficiency, while continuing to improve through practice.
For the sampling proportions of simulation data, real-robot data, video and other sources, we already let the model to 'decide' through an optimization process that selects suitable ratios. This automated decision mechanism is still young and may not be the final optimum, but we are gradually incorporating the idea.
Over the longer term, humans may no longer specify the required data formats or allocation ratios. Machines may determine them through repeated iteration aimed at maximizing performance. We are trying to bring that principle into the entire data system.
ReSpark: Your company has introduced portable equipment for collecting data during everyday activities, with large-scale rollout planned for the first half of this year. Compared with existing solutions, what distinguishes it? How does it improve collection and analysis, and what does it cost?
Cewu Lu: Each existing collection device has advantages. From our own perspective, I think our solution is comparatively optimal, although that is my personal assessment and other teams may disagree. To discuss the differences, we can begin with teleoperated robot arms.
One drawback of teleoperation is difficult for us to accept: it interrupts people's normal lives. Historically, collecting data for general intelligence has usually not greatly disrupted everyday activity. When you publish a blog or review food, you are meeting your own needs while incidentally generating training text for a large model. Autonomous driving is similar: data can be collected while people drive, without taking them away from their regular work. Once people must leave normal activities to perform dedicated annotation, scaling becomes difficult.
With teleoperation, someone has to go to a laboratory and control an arm, affecting their life and work. Operating through an arm can also feel like scratching an itch through a boot: you are not directly touching the object or feeling its force feedback. It commonly requires a fixed site, purchasing a robot and employing someone specifically to collect data, all at considerable cost. As a result, most teleoperation-based collection can only remain in laboratories and is unlikely to capture data from real deployment settings.
Our alternative is an exoskeleton that accompanies people through their work. First, it scales because it does not excessively disrupt ordinary production or daily life. People wear lightweight equipment and can handle objects naturally during real work, substantially reducing the total cost to society. Second, it captures real scenarios: the person interacts with actual objects, rather than performing a simulated laboratory operation. Third, the equipment itself is inexpensive, lowering the overall cost and making wider adoption easier.
Academic research alone may not require a startup to consider social and economic costs in the same way. Real deployment, however, requires solving technical problems within those constraints. Our approach has generated many orders, and the government also hopes we can expand the collection system quickly. Increasing manufacturing capacity still takes time, particularly with the production schedule under pressure around the Lunar New Year.
Motion capture is another approach, but, like teleoperation, it also lacks real physical feedback in the comparison I am making. For example, while trying to grasp an apple, the virtual hand may penetrate its surface: it appears to hold the apple without genuine contact or force feedback. To me, an operation without any tactile feedback is like the emperor's new clothes—pretending to make clothing when there is none. That absence is uncomfortable. This comparison concerns operating a separate virtual or robot embodiment; capturing a person who directly handles a real apple does not itself remove their physical feedback.
Handheld devices are another promising approach because they are inexpensive and can be used on site. But I see some problems. First, our exoskeleton can mathematically guarantee 100% isomorphism with its corresponding robot: the two have identical mechanical structures, so there is no trouble transferring the data to that robot. A handheld device needs the robot's joint configuration and mechanical parameters to be worked out afterward. A person may perform a movement outside the robot's range without realizing it, producing invalid data that can become harmful training noise. Additional algorithms are needed to filter such examples, introducing further indirect errors. The structural correspondence I am describing is to the robot configuration matched by the device, not a claim that one exoskeleton has the same mechanics as every robot.
Second is cost. A handheld solution costs almost the same as our exoskeleton, but in our comparisons the exoskeleton offers better data quality and closer correspondence to robot operation. Accurate handheld localization often relies on SLAM or inertial measurement units, and its present accuracy is not ideal. Better technology may improve it in the future. We have also built handheld devices and are addressing these issues internally. We remain open to multiple approaches. For now, our efficiency comparisons favor the exoskeleton in compatibility with ordinary work, cost and accuracy.
ReSpark: PI uses substantial amounts of Open X-Embodiment data in pretraining. Your team is an important contributor to this dataset, described here as the world’s largest public robot dataset. Do you use it in your own pretraining? How valuable is this open data to companies?
Cewu Lu: I first want to emphasize that data quality matters more than quantity. I have said this repeatedly, and I think people recognize the point. Much existing data is large in quantity but sometimes collected simply to increase the count, without sufficient quality. In our experiments, removing low-quality examples and retaining only the good ones actually improved performance. More data can also mean more noise, so we must consider both sides.
Open data at least allows researchers to experiment today. In the longer term, though, we must recognize that the required scale will be much larger. That is why we pursue collection that accompanies normal activity and can scale. The exoskeleton is one feasible method; others may emerge. Unless collection runs alongside production and everyday life, too few people participate, and the resulting data is insufficient for industrial needs.
Autonomous vehicles, for example, collect enormous amounts of data, even though their decisions largely concern movement on a two-dimensional road surface. We need a scalable route that involves frontline workers across China and ultimately the world. Only then can we obtain enough data. Compared with driving, our range of physical situations is broader and our data requirements greater. We must expand the volume while reducing redundancy and noise. We have done extensive work, including building a foundation model of the physical world, to compress the uninformative parts of that space.
Personally, I am certain that today's open datasets are insufficient for industrial applications. They may support academic demonstrations, but fall far short of stable, robust deployment. We pursue two complementary routes. First, we jointly consider social, economic, cost and technical factors to develop new scalable collection methods. Second, we reduce complexity at the model level and introduce more useful information, including our physical world model for manipulation. Only by compressing both ends of the data problem in this way can we hope to industrialize the technology.
ReSpark: Both major embodied-intelligence companies and startups are building their own datasets. How do you view this?
Cewu Lu: A dataset should not be considered in isolation, and we should not simply assume whoever has the most money wins. That may have been closer to the situation in the previous generation of large models. Now and in the coming phase, data, models, evaluation and systems have to be integrated. Model research, system evaluation and collection are inseparable.
People talk about annotating vast quantities of data, but is that data useful? Unless its effectiveness can be demonstrated, the number of examples is not proportional to the information they contain. I see two phases: before understanding converges, and after it converges. In the first, the key is insight into which data benefits the model, how to assess its value in industrial settings, and whether a meaningful evaluation system can be built. This creates a flywheel in which data, models and applications improve together. Can the collected data improve the model? Can that improvement create value in deployment and then guide better collection? These elements must reach a balance.
Many people look only at the number of collected examples and overlook that understanding. At this stage, we need to know internally which data helps the model and which model suits a particular application. Competition becomes primarily about resources and funding only after that understanding converges. In my comparison, language-model developers began with clearer ideas of data formats and useful-data standards, so competition quickly became about spending. Embodied intelligence—or “resource intelligence,” another term used but not defined in the original transcript—first requires insight: a data flywheel must drive a model flywheel and then an application flywheel, creating a functioning cycle rather than an idle system accumulating data that only appears useful.
We have built an integrated internal system linking data, models and evaluation. Every collected data item is checked for effectiveness, and feedback guides iteration. We identify saturated data—for example, after collecting thousands of door-opening actions in one scenario, further repetition may be unnecessary—and use that information to guide subsequent collection and model development. This helps us balance understanding with resources and achieve real industrial deployment.
ReSpark: Your team has accumulated considerable expertise in force control, which we see as the most extensive accumulation among current embodied-intelligence companies. You recently released a scalable system for collecting force-related data, accurately recording force, pose and motion trajectories during manipulation. An imitation-learning algorithm then uses diffusion models to predict force and position actions throughout 3D space directly from 3D vision. It already works well on peeling fruit and vegetables. Could this greatly expand robots' manipulation capabilities? What have you achieved so far?
Cewu Lu: Let us step back and consider force. We do not simply start from the belief that force is important; we analyze what the task requires. Embodied-intelligence tasks can be divided into two categories for this discussion. One is free-space motion without contact, where contact forces need not be considered. The other includes most high-value manipulation tasks, which involve contact with objects. Force is then unavoidable. It may not be an explicit variable in the task description, but it determines whether the task can be completed and how well.
Take opening a bottle cap. It looks simple, and you might imagine completing it by moving slowly without explicitly controlling force, but the result is poor if the required force is not applied. Similarly, a robot wiping a table can follow the same trajectory yet produce very different results if it does not apply suitable force. Pressing down with sufficient pressure and the appropriate directional force improves the outcome. This helps explain why contact tasks such as shaving are hard to perform satisfactorily if force is ignored. Physically, opening a cap always involves force; the distinction here is whether the system explicitly accounts for it.
These force patterns can generalize and remain robust. Opening a water bottle, a medicine bottle, or operating a gas stove or tap involves shared mechanical patterns. Roughly locating the object and applying an appropriate force pattern can often achieve the task. That is why we have studied force models for so long. With our large collection of relevant data, training a force-aware model is a very different challenge from training one that relies only on trajectories. It requires more than straightforward imitation: we need a complete model of the force space.
Second, to scale collection, we initially proposed a handheld solution and later upgraded it to an exoskeleton—combining the handheld component with an external mechanism—to capture and use force data more comprehensively. We have already accumulated substantial mechanical-interaction data, but recognize the need to increase its scale further. That remains an ongoing priority.
Third, imagine operating solely through trajectories without accounting for force. Can you complete many tasks well? Try teleoperation using only the motion trajectory, and the results can be very poor. Shaving through that setup, for example, would be unacceptable: without force feedback, you might cut the person or fail to shave properly.
In some of our agricultural applications, we have developed equipment to peel fruit and vegetables. Without accounting for force, the result is ragged, as though a dog had chewed it, or the process becomes extraordinarily slow. It is like surgery, controlling every millimeter and millisecond so carefully that a normally quick task could take hours. Opening oysters, turning a key or opening a door also becomes much less effective if the system relies only on position data. Trying such operations reveals how difficult they are without force feedback: a crucial dimension is missing.
As an analogy, some equations appeared to have no solution until complex numbers were introduced. Adding force as a dimension similarly makes many manipulation problems tractable. I consider it an essential route and a factor we cannot avoid.
ReSpark: Earlier manipulation research mainly focused on the categories of objects and scenes. The trend now is to combine additional information, such as object attributes and affordances, to support stronger logical reasoning. Noematrix has worked at the frontier of this direction, accumulating knowledge that improves robot understanding and reasoning.
Cewu Lu: This also concerns an embodied chain of thought: why do we perform an action in a particular way, and how do people identify the crucial reasoning steps? Breaking the process down reveals several core elements: attributes, affordances or controllability, and functionality. A shoe, for example, may suggest some uses of a hammer because we understand its potential function. A rusty knife may seem less suitable or controllable, making us reluctant to use it. These properties form our basic perception of the world. In this discussion, affordance refers to the actions an object permits, rather than being identical to controllability in the formal control-theory sense.
We have done extensive work here and, judging by the published record, were among the first teams working in this direction. The technology is not exclusive to robotics, but we have made leading contributions in vision by identifying and reasoning about these attributes. Reasoning is central: after extracting basic events and features, how do we use them to understand the whole situation? That is one of our core challenges.
We have also integrated this work systematically. Our system can detect and verify all these affordance-related indicators. In many settings, on certain tasks, we find that our approach performs better than the GPT models we compared with. This is because we do not rely only on large amounts of data; we also incorporate chain-of-thought reasoning, or CoT.
For a simple example, rust on an object may reduce its usefulness, or a full container means we should stop pouring. Reasoning about details like these supports more intelligent decisions. Although end-to-end large models are widely discussed, I believe vast amounts of data alone are insufficient; a chain-of-thought layer is also needed. We have invested considerable work in the data, models and theory, and will release related capabilities over the coming period. This is how we are exploring reasoning and chains of thought in embodied-intelligence systems.
ReSpark: Will you open-source some of this work to benefit the broader community?
Cewu Lu: Yes, and I strongly agree with that idea. Startups should share results that address common problems so that we can develop together and help one another. I am willing to share our views and experience because this supports other startups and the industry as a whole. I believe this will become a very large industry requiring many companies to build it together. We have always maintained an open attitude and have released substantial amounts of data and code. In that spirit, sharing can help all of us move the field forward.
ReSpark: You recently introduced an embodied foundation-model architecture based on RISE and CAGE. You report that its measured metrics exceed the widely used Diffusion Policy baseline, with stronger generalization across objects, backgrounds, robot workspaces, viewpoints and embodiments. What problems does it mainly solve? Do you expect more companies to adopt it?
Cewu Lu: Our approach handles data particularly well and adapts reliably to changes in objects, backgrounds and camera views. That stability is a crucial advantage. In our policy evaluations, the metrics have exceeded the mainstream methods we compared with. We have released the relevant code and models so that anyone can verify the results. I believe more people will use those open resources to bring further innovation to the field.
We are also exploring end-to-end trajectory planning and studying its generalization in depth. Although the research models are public, we may not yet be able to open-source the product models. In our evaluation, the academic models already show better performance than Diffusion Policy and related baselines, supporting our view of the underlying approach. Our support for 3D vision is another strength. Many methods face heavy computation and high cost in processing point clouds; we have completely solved that problem in our system. Editorial note: RISE and CAGE themselves use diffusion-based action prediction. These comparisons concern their perception and policy architectures against tested baselines, rather than replacing diffusion with an entirely different action-generation family.
ReSpark: How is Noematrix Brain progressing? Its rapid-deployment mode is described as learning a single task from a few dozen demonstrations, without introducing a pretrained model, while generalizing across camera views and object positions.
Cewu Lu: There are three layers to consider. The first is the foundation model, the system's core. We have built a model that understands the world, incorporating force as I described earlier. Like the brain's engine, it determines whether the system can understand and execute tasks. The second is the data flywheel. The format of collected data must be closely connected to the model, allowing us to verify usefulness quickly and use the feedback to guide collection. Finally, the general model provides application interfaces above these foundations. As you said, simple demonstrations can teach it to do something well, although how simple the demonstration can be still depends on task complexity. We cannot claim that every task can be taught through a simple demonstration or release-type operation; the exact meaning of 'release-type' is not specified in the original account.
We also introduce the idea of a robot prompter: a prompt or a guiding operation. Whether instruction uses language, physical demonstration or another method, the aim is to minimize the cost of teaching and communication. Like a master teaching an apprentice, the simpler and more direct the process, the lower the cost. This makes actual deployment faster. When a new assistant is intelligent, they learn quickly; when their underlying ability is inadequate, instruction becomes much more expensive.
Ultimately, we have two tasks: make the system intelligent enough and make interaction friendly enough that communication costs remain low. A person can be intelligent yet hard to communicate with, reducing overall efficiency. We will accelerate our work here. The complete system combines a foundation model, an efficient data flywheel, and low-cost deployment through simple demonstrations. This entire system will ensure that our real-world applications progress quickly and reliably.
At present, many indicators cannot be quantified because standards are missing. Unlike conventional large models, even after collecting substantial data, we can only make approximate judgments about whether the system's basic capabilities are sufficiently intelligent. We consider three aspects: its ability to learn from and process data; its underlying intelligence; and its ability to communicate with people at low cost.
Low communication cost is crucial to rapid adoption. There are two deployment costs to consider. First, we should not need large numbers of engineers and scientists to implement a project, because those resources are limited. Second, the people helping with deployment should be able to become robot teachers. We want every ordinary worker to be a teacher, without necessarily relying on specialist scientists. Someone should be able to buy a robot, follow the manual and teach it easily, substantially reducing cost and supporting expansion.
That is why we emphasize this point. Requiring large amounts of expensive labor for each project creates a bottleneck. We want anyone to be able to teach a robot easily, so that adoption and deployment can scale.
ReSpark: Noematrix is among the leading companies demonstrating long-horizon tasks on real robots. How do you make those tasks work? You also recently introduced a new framework for learning long-horizon manipulation concepts from a scalable world model. Could you share its underlying idea?
Cewu Lu: I see three main problems. First is understanding and decomposing the goal: breaking a large objective into executable steps. Large models already perform reasonably well here in research, and major steps in a product task are usually clear, so I will not dwell on that layer.
Second is monitoring and judging intermediate states, which is very difficult. During execution, a robot may not realize that it has failed or know which step it has reached. We therefore propose a deeper chain of thought, the causal chain mentioned earlier, integrating attributes, affordances and functionality. It helps the system judge the state of execution. When pouring water, for example, a system without feedback might keep pouring after the container is full. With the causal chain, it recognizes that the water has reached the required level and stops. This supports more accurate state assessment instead of depending solely on inefficient statistical conclusions from data.
Third is the stability of low-level skills. A five-minute demonstration may look impressive, but a long-horizon task can involve a dozen or more skill modules. If each succeeds only 90% of the time, errors become much more likely across the chain. We have worked extensively on precise causal feedback so that skills connect correctly. When opening a door, the system must judge whether it has reached the right angle and use force feedback to avoid pulling the door off. During insertion, vision alone may not reveal whether the part is seated, whereas small changes in lateral forces can help establish completion. This causal-network mechanism substantially improves robustness.
In our actual tests, we performed tens of thousands of demonstrations over consecutive days in settings such as clothes folding and firefighting operations, with almost no failures. Our goal is to connect goal decomposition, state feedback and stable skill transitions throughout the chain, maintaining accuracy and robustness on long-horizon tasks. That is the basic direction of our work.
ReSpark: In addition to single-arm and dual-arm systems, how is adaptation to dexterous hands progressing? How well can your system transfer across hardware embodiments?
Cewu Lu: First, on hardware adaptation: I think the academic community has gradually moved toward the view that this is no longer a major problem. Our experiments show that it is no longer a barrier in the cases we have studied. We have an interesting paper under review at a Nature-family journal. Using a model trained for a two-finger gripper, we achieved five-finger grasping with only a small amount of adjustment.
A fundamental reason is that much of the information needed for manipulation concerns understanding the world and the task; differences in mechanical parameters account for a smaller portion. Transfer is therefore not difficult using our method. As an analogy, imagine a person's mind entering a chimpanzee's body. The body changes, but the intelligence can still carry out complex work. This illustrates why I regard intelligence as central and mechanical differences as less significant. Looking at individual differences, philosophical considerations and the underlying reasoning, I do not see hardware adaptation as a major obstacle. Our experiments support that view. Hardware adaptation is even less of a problem at the product level because we emphasize adaptive design, without substantially increasing industrial cost.
Dexterous hands are a promising research direction, and we have worked extensively on them, including combining touch with vision. Studying them is academically necessary, but many researchers do not fully understand the core problem: in-hand manipulation, or adjusting the object within the hand using feedback. If you initially hold a hammer poorly, you can shift it into the correct position. You can similarly adjust scissors from an unsuitable grip until you can use them accurately. The value is increased precision and flexibility beyond simple two-finger operation. Our laboratory focuses on these questions and will publish related work soon.
For commercialization of dexterous hands, we consider three factors. First is stability. A product must operate continuously for months under intensive use and complex conditions, meeting industrial standards. Does a dexterous hand meet that requirement? I will not judge particular metrics here, but invite people to consider the question. Second is cost, especially return on investment and comparison with human labor. If the product costs tens of thousands of renminbi, cost control must be strict enough to offer comparable or better value than labor. Third is necessity: does it solve a problem for which the technology is indispensable? If a simple two-finger gripper can already complete the task, a more complex hand has less value. We want a simple, stable solution for tasks that two fingers cannot easily perform, with stronger ecosystem support.
The intersection of those three conditions determines the future industrial opportunity and when the technology matures. We can build an assessment system around them, specify the indicators, and judge when commercialization standards have been reached.
There is also AI dexterity. More joints and complex operations require greater adaptability and learning capacity from AI. The complete system becomes practical only when both its intelligence and its hardware meet the requirements.
Regardless of industrial developments, I always encourage researchers and my students to study dexterous hands. My principle is that if two fingers can easily complete a task, there is no need to study a more complex hand for it. Dexterous-hand technology matters when two fingers cannot do the job or a simple method cannot meet the requirements. I often tell students: 'Do not study dexterous hands for things two fingers can do. Study them for what two fingers cannot do.'
ReSpark: You recently published a paper in Nature Communications on a joint visual-tactile learning framework. It combines a vision system with a stretchable tactile glove to reconstruct rigid and deformable objects accurately. Many companies are now exploring this combination of vision and touch. What specifically distinguishes your approach?
Cewu Lu: This is part of our long-term research preparation. Personally, I do not think products in the near term will necessarily use the most advanced techniques, but we continuously prepare technology for future products. I believe joint learning from touch and vision will substantially improve world modeling, increasing accuracy, stability and functional capability.
A central question in collaborative development from hardware through AI is whether we research them jointly from the outset or develop them separately and combine them later. Several years ago, our team began working closely with sensor specialists to refine the equipment together. Sensor design can then be optimized for the AI, producing an integrated end-to-end solution. In our experience, this significantly exceeds separately developed combinations in performance and stability.
I think it will take time before these advanced techniques play their full industrial role. Many products do not yet require extremely precise modeling. In more demanding settings, especially with dexterous hands—for example, accurately picking up scissors or handling parcels—the advantages will become apparent. The exact industrial value is still being discussed. Scientifically, however, if a dexterous hand can carry out difficult tasks like these, it demonstrates considerable potential and means that our research will profoundly influence future product deployment.
ReSpark: How did you decide to focus on a software-only model? Do you think a company providing only embodied-intelligence software has a promising future?
Cewu Lu: We can consider both the eventual outcome and the path toward it. In the long run, a software ecosystem is the most economical arrangement: one brain empowers different kinds of hardware, with each doing its part. That is flexible and economical.
Along the way, however, a purely software-focused business in China faces challenges. We prefer a team centered on AI software that also has strong hardware capabilities. This fits our origins: having emerged from Flexiv, we naturally possess substantial hardware R&D expertise.
Some customers need an integrated hardware-and-software solution. Our strength is starting from their needs and quickly assembling a complete hardware system. A Fortune Global 500 company once asked us to integrate different components—selecting the best arm and the most outstanding mobile platform, for example—into a prototype rapidly, with a path to mass production. We have a mature methodology for manufacturing at scale. Even when we do not perform production ourselves, we can provide the full solution.
Our model is flexible. If a customer has mature hardware, we load our software onto it. If not, we can assemble and deliver the whole system. Meeting different needs, from integrated delivery to software support alone, is the key. This capability lets us respond quickly to customers' requirements at multiple levels.
I will not discuss a specific commercial route, since every route is still being explored. But I like to approach the question in terms of economic efficiency, minimum cost and the interests of ordinary people. If we have to use one fixed hardware form, it cannot suit every application. We must consider cost as well as effectiveness.
I expect different applications to use the most suitable robot configuration. For some cleaning tasks, a single arm on a wheeled base is sufficient; why insist on two arms if they offer little extra benefit but higher cost? The market can support many robot types configured for different settings. It is like assembling a computer, selecting a CPU, graphics card, GPU and other components to fit the budget. If we offered only complete-machine solutions, the market would instead become very small.
Our ecosystem should match hardware, software and applications flexibly according to economic and operational efficiency, meeting different users' requirements for cost and performance as effectively as possible.
Even with software at the core, strong hardware capability is necessary. When a customer needs physical hardware, we must provide it rapidly. We have close relationships with hardware manufacturers and know their respective strengths and weaknesses, allowing us to connect them quickly through stable interfaces.
When a customer needs a robot, we can combine hardware modules from different companies. For example, we might purchase a batch of 'TAB' components, as the original account calls them, and a mobile platform, then integrate them into a complete robot. The account does not explain the TAB abbreviation. We have developed many templates for different combinations and embodiments. Close cooperation with manufacturers allows the partners to work smoothly together and means we do not have to worry about last-minute problems.
Assembling modules may sound easy, but rapidly connecting hardware through stable interfaces while preserving overall performance involves many details and coordination challenges. Hardware, software and interface developers must work closely together to achieve efficient deployment.
ReSpark: How do you view commercial deployment?
Cewu Lu: I see embodied intelligence as an industry where technology enables the market, rather than one driven simply by existing market requirements. Demand may be almost unlimited, including situations we have not yet imagined. Many needs remain unmet because our skills are not sufficiently advanced. Once those skills develop and the technology reaches the necessary level, it can quickly fill the gaps.
Deployment therefore depends strongly on advances in technology and research. Our work on AnyGrasp is a good example: general grasping means handling many different grasping situations. Some results have already been deployed, both by us and by other teams. Before the breakthrough, many applications were infeasible and could not create revenue. Once the technology emerged, it solved part of those problems and created value. Fundamentally, the change came from a technical breakthrough.
Internally, we maintain a timetable forecasting when particular skills will become relatively mature or general enough to use. Engineering support and customization can then bridge the remaining gap. A capability at roughly 90% maturity, for example, may be made fully deployable in a particular industrial setting through supplementary measures. We make qualitative judgments about maturity and consider potential applications in advance. Commercial confidentiality prevents me from sharing the details, but the broad idea is that some skills will reach full maturity and others about 90%, with on-site engineering support unlocking successive groups of applications. These percentages describe our planning and deployment reasoning, rather than a guarantee of 100% success in every situation.
People also ask when embodied intelligence will have its 'GPT moment.' It cannot happen all at once because our systems are much more complex than language models. Applications will unlock in stages: one group of scenarios and skills this year, more next year, and another group the year after. We may see successive, industry-specific GPT moments, driven by skills maturing one after another. With AnyGrasp, for example, reaching a high level of industrial deployability can open a whole group of applications.