Z Potentials | Exclusive Interview with Dexmate: Redefining Dexterous Manipulation with MIT & UCSD Roots, Software-Hardware Synergy, and a Data Flywheel Strategy
Z Potentials invited Tao Chen and Yuzhe Qin, co-founders of Dexmate, to give a talk.
In this interview, we sit down with Tao Chen and Yuzhe Qin, co-founders of Dexmate, a startup focused on AI-powered robotic hands. With deep research backgrounds at leading labs from Shanghai Jiao Tong University, MIT, Carnegie Mellon University (CMU), and the University of California, San Diego (UCSD), the duo brings both technical expertise and a clear understanding of industry needs.
Dexmate is developing next-generation dexterous robotic hands by tightly integrating hardware design, intelligent control, and data-driven learning. Their core strategy—what they call a “data flywheel”—combines simulation data with real-world datasets to significantly boost the adaptability and generalization of their system. The company’s AI-first approach aims to enable robotic hands to perform a wide range of tasks, from manipulating tens of thousands of different objects to interacting with various door types. All computation and control modules are embedded directly into the robot, achieving a truly plug-and-play solution. This design philosophy—requiring no special integration or modification from users—lowers deployment barriers and makes robotic dexterity widely accessible across industries.
Unlike many robotics companies that pursue human-like form factors, Dexmate doesn’t aim to replicate the human hand visually. Instead, they focus on functionality and task efficiency. Technically, they adopt a co-design approach, aligning mechanical structures, sensor arrays, control policies, and AI models from the ground up. The goal is simple: maximize real-world performance.
Join us as we dive into Dexmate’s origin story—enjoy! :)
“Our robotic hand isn’t about mimicking the human hand’s appearance—what matters is replicating its capabilities.”
“An ideal dexterous hand should be like a human one: a rigid skeletal core wrapped in soft material. This hybrid design ensures both strong gripping force and high load capacity, while soft surfaces improve friction and tactile precision.”
“We emphasize hardware-software co-design. From day one, we engineered the hardware with downstream AI model training and deployment in mind—a key differentiator for us.”
“At the algorithm level, we’re strong advocates of sim-to-real strategies, blending synthetic and real-world data for more robust task execution.”
“We aim to leverage existing datasets to accelerate the generation of new training data, allowing the data flywheel to spin at an exponential rate.”
“Building robots is like running a marathon. You need more than just short bursts of brilliance—you need long-term stamina and persistence. Robotics is a deeply interdisciplinary field that requires patience and resilience.”
01 From Top Labs at MIT and UCSD: The Rise of a New Force in AI Robotics
ZP: Can you briefly introduce yourselves and your past experiences?
Tao Chen(Left), Yuzhe Qin(Right)
Tao Chen: I’ve been independent since I was young—I started boarding school early. I’ve always held myself to high standards, both personally and professionally. I was into STEM from a very early age, competing in all kinds of science contests. I got into Shanghai Jiao Tong University’s mechanical engineering program through a talent-based admission track. But later, I realized that the true core of robotics lies in software and algorithms. So during my senior year, I taught myself computer science and pivoted to AI.
After college, I joined a startup working on humanoid robots, where I worked on SLAM and reinforcement learning. Then I went to CMU, where I focused on applying reinforcement learning across different types of robots. I also worked on a project with Meta that aimed to lower the entry barrier for robotic development.
At MIT, I began my PhD research on quadruped robots, and by 2019 we had already trained a robot to climb over obstacles using reinforcement learning. By the end of 2020, I shifted to dexterous manipulation, which I believe holds greater commercial value—and is also one of the most technically challenging problems in robotics. I’ve spent the past four years deeply involved in this area and have published extensively.
Yuzhe Qin: Hearing Tao’s experience of growing up in boarding school really struck a chord with me. I was also raised in a city different from where my parents lived, so I had to become independent early on. Our paths are remarkably similar. We both entered the Mechanical Engineering program at Shanghai Jiao Tong University through independent admission, and later transitioned into AI and dexterous manipulation research.
I’m two years younger than Tao at Jiao Tong. After graduating, I moved to the U.S. At the time, seeing how fast AI was evolving, I realized robotics couldn’t be solved purely from a mechanical perspective—it had to be an AI problem. I was fortunate to become the first student working on robotics in Professor Hao Su’s lab at UCSD. During my PhD, I was also Professor Xiaolong Wang’s first doctoral student. Like Tao, maybe because we were early members of our labs, we learned to work independently and take initiative.
Personally, I’ve always had a weak constitution growing up. I started running in high school to improve my health and stuck with it for over a decade. I’ve competed in more than 50 races, including over a dozen marathons. I may look thin, but back in Jiao Tong, I was the captain of the running club. I often think building robots is a lot like running marathons—it’s not about short bursts of speed, but long-term perseverance. Robotics is an interdisciplinary challenge. It’s never a quick win. It takes grit, stamina, and sustained focus.
ZP: Why did you choose this particular moment to start a company?
Tao Chen: Now is a perfect time. AI technology has matured enough that it can be applied to training robots for much more complex tasks. This judgment is based on my personal experience over the years. I’ve been deeply involved in AI-driven robot control, both in legged locomotion and dexterous manipulation.
Through my work in reinforcement learning, I’ve witnessed firsthand how quickly things are advancing. Take quadruped robots—when we started in 2019, training them to walk over obstacles was incredibly difficult. But just a year later, the problem was largely solved. The same pattern is playing out with dexterous hands. Four years ago, only a handful of researchers worldwide were working on this, and the challenges were immense. But with steady progress, the field is clearly reaching a turning point.
Yuzhe Qin: For the robotics industry, this is the right moment for the commercial world to take over. In most deep tech domains, academia paves the way by exploring fundamental questions. Once the methodologies mature, it’s up to industry to scale and refine them into real products. That’s exactly where robotics is today—ready to transition from academic prototypes to real-world, engineered systems.
And when it comes to starting a company, timing and the right partner are everything. I would never have done it alone. But when Tao approached me, I didn’t hesitate. We’ve known each other for years. He’s both deeply committed and incredibly reliable, so I knew I could trust him—and the mission.
ZP: How do your research focuses in dexterous hands differ?
Tao Chen: My work leans more toward large-scale simulation training using reinforcement learning. I focus on training dexterous hands to perform complex manipulation tasks in simulators, then transferring those trained policies to physical robots. For instance, we work on how to make a robotic hand grasp and manipulate thousands of different objects using minimal sensing—sometimes just a single RGB image from a camera. The core idea is to solve high-difficulty control problems using the simplest, most cost-effective system, while ensuring the solution is robust.
Yuzhe Qin: Our entry points are quite different. Tao’s background is in controlling quadruped robots, so he naturally approaches manipulation from the control systems side. I started from a computer vision background, so my perspective is rooted in perception—specifically, how to gather data from the human world to support manipulation learning.
Learning from human demonstrations, including video, motion capture, and kinesthetic teaching, has become a hot topic recently. But we started this line of work as early as 2020. Since dexterous hands are inherently anthropomorphic, it makes sense to draw on human data. Unlike parallel-jaw grippers, which don’t map well to human anatomy, dexterous hands can naturally benefit from learning strategies based on how humans interact with objects. After a few years, it turns out most researchers are now converging in similar directions.
ZP: What are the major research challenges in the field of dexterous hands right now?
Tao Chen: The first main thread is large-scale training in simulation—this includes reinforcement learning as well as classical control algorithms like trajectory optimization. All of these approaches begin with intensive training in simulated environments, and then the trained models are transferred onto real hardware. Another major path involves collecting real-world manipulation data. For instance, we might use devices like the Vision Pro, Quest headsets, or specialized gloves to control a real dexterous hand, recording the manipulation data in the process. This data can then be used for imitation learning, helping train neural networks.
Then there’s the method we just mentioned—learning from videos. That approach is attracting more and more attention. But it’s inherently more uncertain, and in many ways it’s still a frontier topic—less technically mature than the first two directions.
02 Breaking Through with Software-Hardware Synergy and the Data Flywheel
ZP: How is the product positioning of the Dexmate dexterous hand shaping up?
Tao Chen: What matters most in a dexterous hand is functionality, not form. We care primarily about whether it can handle tasks with the flexibility of a human hand—whether it can deal with objects of various shapes, or manage complex contact interactions. So we believe a dexterous hand should be defined from the perspective of the tasks it can accomplish.
A lot of people equate dexterous hands with human hands, but in truth, it doesn’t have to follow the five-finger model. That form is certainly welcome, but it’s not the essential benchmark. What really counts is how dexterous the mechanism is.
ZP: When it comes to overall performance, are there any trade-offs—say, flexibility versus rigidity?
Tao Chen: A good dexterous hand should have two core traits: high degrees of freedom, and a well-balanced structure between stiffness and softness. Ideally, it should be like the human hand—rigid bones on the inside, but a soft outer layer. That design allows for strong grip strength and load-bearing, while the soft surface materials provide better friction and a more intuitive sense of touch.
The soft exterior also brings a hidden advantage: it reduces the precision demands on the AI controller. Because the fingertips can deform, they can naturally absorb small control errors. Think about how, in daily life, it’s easier to pick up a plush toy than a glass cup—the toy deforms to increase contact area, making the grasp much more forgiving.
Yuzhe Qin: Regarding degrees of freedom, what we emphasize is effective freedom—not just counting how many motors or joints there are. For example, if a finger has three joints but can only bend downward, its real usefulness is quite limited. But if it can also move side to side, then that’s an entirely different—and far more effective—degree of freedom. So we prefer to define dexterity by what the hand can do, rather than how it looks or what parts it has.
In real-world applications, there are also many practical issues to consider. Motors might overheat during long-term operation—sometimes we’re talking about robots working continuously for over ten hours. And in a household setting, the robot may come into contact with water or food waste, so waterproofing becomes essential. We also need tactile feedback so the robot can control its grip—so it doesn’t crush a bowl or break an egg.
ZP: What performance benchmarks have your products achieved—like precision in effective degrees of freedom, load capacity, response speed, and so on?
Yuzhe Qin: As an AI company, aside from the hardware benchmarks you just mentioned, we pay close attention to AI-related metrics. The dexterous hand we’re developing actually surpasses the human hand in key areas—it’s even more flexible in some respects. We’ve also equipped it with a suite of environmental sensors to improve its adaptability to dynamic situations during manipulation tasks.
We’re moving forward on two tracks simultaneously: independent development and platform adaptation. On one hand, we focus most of our effort on training AI models, so we make use of top-tier dexterous hands already available on the market—our AI models are platform-agnostic and compatible across devices. On the other hand, for certain specialized scenarios, we do build our own hands.
Algorithmically, we place a strong emphasis on what we call “virtual-real fusion.” Different companies have different focuses—Physical Intelligence, for instance, prioritizes the quality of real-world data, while NVIDIA emphasizes the scale advantages of simulation data. We believe both approaches have their strengths; there’s no need to choose one over the other. Our team has built up deep expertise in both areas, and we’re currently exploring technical strategies to combine real and simulated data in a seamless, task-oriented manner.
And we care deeply about software-hardware synergy. Even in the early stages of hardware design, we fully consider the downstream demands of AI development and model deployment. That’s a major differentiator for us—one of the things that makes our approach unique.
Tao Chen: We see robotics as a complete systems engineering problem—especially when it comes to data acquisition, we focus more on real-world effectiveness than academic debates. Whether it’s simulation data, human video data, or any other form—if it’s useful, we’ll use it. Because a truly powerful robotic foundation model must draw from a fusion of diverse data sources. Our goal is to build a comprehensive toolkit: whenever a certain type of data is needed, the most suitable tool is at hand to capture it.
Take real-world data collection. For instance, our team has done pioneering work. We were the first in the world to use the Vision Pro to control a dexterous hand for manipulation, and later open-sourced the code, which has since been adopted by many companies both in China and abroad. In terms of simulation, we’ve also built up extensive experience, and were among the earliest teams globally to use GPU-based large-scale parallel simulators.
However, collecting data from an AI-centric perspective alone is not enough—there’s a gap at the level of physical machines. To implement real-world functionality, one needs a deep understanding of control systems—basic motion control, force control, even the fundamentals like PID control. These are areas easily overlooked by those with purely AI backgrounds.
Thus, from a practical deployment standpoint, we adopt a pragmatic approach, while maintaining a systematic mindset. On the one hand, we’re building a complete toolchain, capable of mobilizing resources on demand. On the other, we emphasize the integration of hardware and software—optimizing across the system to deliver the best product experience possible.
ZP: What are some key milestones for the company going forward?
Tao Chen: Since we develop both hardware and software in tandem, there are two major milestones ahead. On the hardware side, we aim to create a piece of intelligent hardware that genuinely satisfies our clients. This won’t be a traditional robotic arm that simply executes pre-programmed motions, but an AI-powered smart device capable of performing a wide range of tasks autonomously.
On the software side, our goal is to develop a general-purpose dexterous hand model—one that can operate across scenes and across tasks. Think of it like ChatGPT: while not tailor-made for any single job, it serves as an excellent pre-trained foundation. Once it’s in place, fine-tuning it for a specific task becomes quick and efficient. This is our major algorithmic objective.
ZP: What does the training process for a dexterous hand model look like? What key metrics do you emphasize?
Tao Chen: Data diversity and volume are our two most valued indicators. As for the collection process, we don’t impose limits—if we can think of a task, we’ll try to collect for it. Regarding diversity, we must go beyond single-scene repetition. Take grasping tasks, for example—if all your training data comes from one tabletop, even if the volume is high, the lack of variety renders it less meaningful.
So how do we scale up the data? In the real world, teleoperation is a common method. The key is to make the teleoperation devices simple enough that even untrained users can operate them. At the same time, we aim to improve teaching efficiency—shortening the time it takes to demonstrate a task from a minute to just one or two seconds.
Beyond these system-level optimizations, we’re also pursuing scalability on the simulation side. Our goal is to construct a general framework, not the one that needs custom code for every new task. This way, when we want to add a new task later, it’s just a few lines of code away.
Yuzhe Qin: As for data, our ambition is to build an exponential data engine. Why exponential? Typically, data collected through real-world teleoperation grows linearly with time. But what we’re trying to achieve is the use of existing data to accelerate the generation and collection of future data—so the data flywheel can spin up exponentially. If we continue with traditional, linear collection methods—say, just by using teleoperation—then it will be very hard to achieve a fundamental breakthrough in machine learning within a few years.
ZP: If we draw an analogy with foundation models—after GPT emerged, academic methods quickly converged. Where is dexterous manipulation now, on that timeline?
Tao Chen: Dexterous hands are still a relatively nascent field—real technical development has only happened in the last few years. While we’ve accumulated considerable experience, reaching a ChatGPT-style inflection point hinges on one thing: amassing large-scale data.
Right now, a lot of cutting-edge research remains stuck in academia—but the academic world often suffers from a lack of large, high-quality real-world data. This is where startups have a real advantage. Not only do we master state-of-the-art control technologies for dexterous hands—we’re also embedded in real use scenarios, enabling us to collect real-world data at scale. That’s the only path to the “ChatGPT moment” for robotic manipulation.
ZP: What is the approximate amount of data required?
Tao Chen: Training data in the robotics domain is fundamentally different from that of ChatGPT. Whereas ChatGPT can effortlessly harvest vast and effective datasets from the open web for next-token prediction, robotics is an entirely different beast.
Here, every piece of high-quality demonstration data is uniquely precious. To speak merely of volume, while ignoring quality, is simply misguided in our field. Take reinforcement learning in simulators, for instance—perhaps 90% of the generated data is effectively useless. Reinforcement learning, by its very nature, is a process of stochastic wandering: occasionally it stumbles upon the right signal and then reinforces it, but along the way, it generates torrents of noise.
These days, a single GPU might churn out data by billions within an hour, yet less than 10%, perhaps even under 1%, is truly meaningful. The truth is, no one in the field can give a definitive answer—how much data is “enough,” and what kind, remains an open question.
Yuzhe Qin: Though we do not yet possess enough data to fully crack the code of dexterous hand modeling, we’ve discovered an effective path for accumulating it. Within our current training framework, we’ve already managed to accomplish relatively simple tasks that hold real-world utility across various industries.
This gradual and grounded approach allows us to steadily expand our data reservoir, laying the foundation for more complex use cases. I firmly believe that within the next three to four years, we’ll have gathered enough practical data to truly move the needle. Even if we remain uncertain about the exact quantity required for a universal dexterous hand model, that doesn’t stop us from forging ahead with real-world applications today.
ZP: What are the main hardware challenges?
Tao Chen: From a structural standpoint, the technology itself isn’t particularly complicated—in fact, it’s fairly mature. But bringing it to market requires elevating the performance of every single component, which in turn demands deep collaboration across the ecosystem, especially with supply chain partners.
For example, we were able to build motors for our dexterous hands early on, but they were simply too bulky. If we want real-world deployment, we must push for extreme miniaturization without sacrificing torque or performance. That requires working hand-in-hand with our suppliers to bring every foundational part to a higher standard.
Though this field has seen years of diligent effort, true commercial viability still hinges on meticulous attention to detail.
Yuzhe Qin: Let me add a thought. Building a dexterous robotic hand actually shares surprising parallels with building VR headsets. The core technology isn’t the real hurdle—the true challenge lies in refining every engineering detail to the point where the product is not only functional, but so user-friendly that people are willing to pay for it.
Take thermal management, for instance. As VR headsets get smaller, heat dissipation becomes a major concern. The same goes for dexterous hands: once you pack motors densely into a compact form, keeping them cool becomes a significant issue.
This leads to inevitable trade-offs—cutting ventilation holes may improve heat dissipation but could compromise water resistance. In the end, the challenge is how to harmonize all these technical demands into a coherent and elegant engineering solution.
ZP: Will differences in hand design or sensor configuration affect the transferability of our models?
Tao Chen: The reinforcement learning algorithm we’re using is actually quite general and doesn’t rely heavily on the specific structure of a robotic hand. A better design might help us solve tasks more efficiently, but fundamentally, the algorithm works across different configurations.
At the end of the day, all algorithms — whether reinforcement learning or neural networks — are essentially methods of search, seeking out solutions to problems. A well-designed hand makes that search easier. During my doctoral research with Yuzhe Qin, we worked with at least five different dexterous hands, but the underlying logic of the algorithms remained much the same.
Today’s algorithms are already general enough to be extended across different hands. However, if a model is trained on one particular hand, it can’t yet be directly transferred to another. Still, I believe this problem will be solved soon. For example, once every hand is capable of handling five different tasks, it becomes much easier to train a model that can control multiple types of robotic hands. The real challenge isn’t model transfer across hands, but teaching a single hand to perform a wide variety of tasks.
03: The “Plug-and-Play” Revolution — AI-Powered Dexterous Hands as Versatile Assistants from Kitchen to Factory
ZP: What’s the design philosophy behind Dexmate?
Tao Chen: Our product is guided by two core principles. First, we pursue a design that is driven by AI, optimizing both hardware and software in tandem. Second, we place a strong emphasis on real-world performance and robustness. We’re not just making a flashy demo or a marketing video — we’re building dexterous hands that can truly perform complex tasks. If it’s grasping, the hand needs to handle tens of thousands of object types. If it’s opening doors, it must adapt to all sorts of mechanisms.
Our vision is to make robots truly plug-and-play. All the necessary computational resources are built into the system, so users don’t need to make special adaptations. The robot can step right in and replace manual labor without requiring businesses to change their environments. That’s the only way we can lower the barrier to adoption and make robots useful across a wide range of industries.
ZP: Who are our customers, and what are they using the dexterous hand for?
Tao Chen: At the moment, we’re mainly serving industrial clients — manufacturers, warehouses, and the food service sector. Even though we’re technically selling robotic systems, what customers are really paying for is the robot’s ability to solve real problems.
Most of these clients are still relying on manual labor and haven’t achieved automation yet. One kitchen client, for example, wants the dexterous hand to stir-fry, flip a wok, and add seasoning. Traditional suction cups or grippers simply won’t do — take wok flipping, for instance. A simple gripper can’t hold the pan securely by the handle. A human hand succeeds because it can wrap the entire palm around the handle. That’s where the dexterous hand shines — it mimics human flexibility and adaptability.
Another example is cleaning work. It’s not just about picking up and putting down towels; the hand must operate spray bottles and various cleaning tools. Such complex actions are beyond the reach of basic grippers. So the great advantage of dexterous hands lies in their high degrees of freedom — the ability to handle different objects and tools. That said, the industry is still young, and many clients are still in the experimental phase, testing what’s possible.
ZP: When can we expect to see the hardware product?
Tao Chen: We’ve designed a mobile robot equipped with dexterous hands. It has a mobile base and a dual-arm configuration, and it will be released in a few months.
ZP: What are the three most important things for the company in the coming year?
Tao Chen: First is PMF — product-market fit. We have the tools — dexterous hands, robots, and AI training frameworks — but the key is to find the most appropriate application scenarios, the real customer pain points. That’s what determines our ability to survive. And to build a general model with strong transferability, we need vast, diverse data. So finding PMF is also the key to spinning the data flywheel.
Second is the team. Building all this requires an outstanding and well-matched team. We need strong AI talent, hardware experts — but above all, everyone must genuinely love robotics. They need that inner drive to build something truly exceptional.
Third is AI technology. We need to stay abreast of industry developments, but also stay committed to this path of combining simulation with real-world applications. The challenge is boosting generalization and robustness — solving real problems, not just lab demos — and making our technology truly land in the hands of customers.
ZP: What is the company’s long-term vision?
Tao Chen: We’re neither purely a hardware company nor solely a software one—we build robotic products. These products rely on powerful AI capabilities and robust hardware systems. Since we focus on dexterous manipulation, we’re not confined to any single industry. Instead, we aim to integrate data across sectors through robotic dexterity, creating a data flywheel to drive faster growth.
To sum up, we want to build robots that are versatile, nimble, and reliable. Our ultimate goal is quite simple: to give people more time to do what they love, by letting robots take care of tedious, dirty, and exhausting tasks. That’s the future we hope to realize.
Rapid Fire Round
ZP: What was the most exciting development in AI and robotics over the past year?
Yuzhe Qin: The current technical landscape includes optimal control, reinforcement learning, imitation learning, and learning from video. These were once distinct schools of thought. But in the latest wave of robotics, researchers and engineers from different backgrounds have begun to work together. Everyone now sees that no single approach is enough.
Take Boston Dynamics for example: though their roots lie in control theory, they now also embrace AI and reinforcement learning. Meanwhile, AI researchers are realizing that control theory is fundamental to building solid algorithms.
This convergence has led to a shared belief: we must integrate AI algorithms with vision, control, and planning across disciplines. That cross-disciplinary momentum is now gaining real traction.
ZP: How do you view future opportunities in China and the U.S.?
Tao Chen: China has an exceptionally strong supply chain—a key advantage in creating superb product experiences. And even in a worst-case scenario, where the global market splits, both domestic and international markets remain vast on their own. For startups, each side holds immense potential.
ZP: What was the most unexpected discovery in your entrepreneurial journey?
Tao Chen: At the beginning, it was just me, Yuzhe Qin, and another co-founder, all with backgrounds in AI and algorithms. We were young—none of us had yet turned 30—and we weren’t sure we could build a top-tier hardware team.
What surprised us most was that many senior engineers, with 10 or even 20 years of experience, chose to believe in us and join our team. That instantly elevated our capabilities across software and hardware. To be honest, we were extraordinarily lucky in that regard.
ZP: What kind of colleagues are you hoping to attract?
Tao Chen: We especially admire those who genuinely love robotics—who have that self-driven passion and truly believe robots can become intelligent. If someone brings that spirit, we’re excited to work with them. As for technical background, robotics is complex enough that almost any skill set can find its place.
Yuzhe Qin: Let me lighten the mood a bit. On RedNote, I often see people asking, “Is embodied intelligence a reliable career path? Should I choose robotics or autonomous driving?” These questions are everywhere.
In my view, embodied intelligence—robotics—is a cutting-edge field. No one can say with certainty what it will look like in three years.
But whether you’re founding a company or joining one, the most important thing is belief. You have to believe in the possibility of success, or it likely won’t happen. It may sound idealistic, but in emerging fields like this, we’re really looking for people who truly believe in what we’re building.
Disclaimer
Please note that the content of this interview has been edited and approved by Tao Chen and Yuzhe Qin. It reflects the personal views of the interviewees. We encourage readers to engage by leaving comments and sharing their thoughts on this interview. For more information about Dexmate, please explore the official website: https://www.dexmate.ai/.
Z Potentials will continue to feature interviews with entrepreneurs in fields such as artificial intelligence, robotics, and globalization. We warmly invite those of you who are hopeful for the future to join our community to share, learn, and grow with us.
Image source: Tao Chen and Yuzhe Qing