In this interview, Sicheng Yang, founder of Dexcelbot, argues that the real bottleneck in embodied AI is not locomotion but dexterous manipulation. He explains why hands, tactile sensing, and long-life hardware matter more than humanoid form alone in turning embodied systems into useful industrial tools.
You moved from building mobile robotic platforms at Robotics X to focusing almost entirely on manipulation; at what moment did you conclude that hands, not legs, were the structural bottleneck in embodied AI, and what trade-off did that decision permanently lock you into?
I joined Tencent’s Robotics X Lab as an early founding member in 2018. For the first two years, I was primarily focused on building mobile robotic platforms, with the multimodal quadruped robot Max being one of my representative projects. After completing Max, I realized that the technical paradigm for legged robots was gradually converging—it was quickly shifting from a research challenge to an engineering optimization problem. At the time, there were already many quadruped robot companies in the market, but none had truly found a scalable commercial application. Back then, the concept of embodiment wasn’t widely discussed. The reason I shifted my focus to dexterous hands and manipulation at the end of 2019 was quite straightforward: I recognized that the practical application value of manipulation far exceeds that of mobility, because most human production activities rely on “hands” rather than “legs.” For a researcher committed to building more useful robots, moving from “legs” to “hands” was a natural progression. I believe the main challenge in designing a “hand” lies in balancing performance and reliability. Cost is also important, but analyzing the BOM of the core underlying components shows that, in the long run, cost won’t be the fundamental bottleneck.
Apex Hand balances 21 DoF, load capacity, and precision rather than maximizing any single metric; what capability did you knowingly sacrifice to achieve that balance, and when did you hesitate before committing to that constraint?
We consciously sacrificed some palm thickness—the back of the Apex Hand is slightly thicker than a human hand. Our team has developed several generations of dexterous hand products, and years of hands-on experience have given us a clearer understanding of how to define a dexterous hand suited for the era of embodied intelligence. For a human-sized, high-DoF dexterous hand, reliability is the top design priority, and the three-dimensional size of the fingers must also be preserved. However, a slightly thicker palm has almost no impact on core functionalities like grasping and manipulation, aside from a minor aesthetic difference. Therefore, we didn’t hesitate much in making this trade-off.
You chose a linkage-driven architecture over tendon systems despite the latter’s biomimetic appeal; where has that decision already shown structural weakness under sustained industrial stress, and what cost do you accept to maintain robustness?
To clarify, the Apex Hand does not use a linkage-driven architecture; it employs a tendon-driven hybrid design. Linkage-driven hands are essentially purely rigid transmissions. We believe they are more suitable for low-DoF dexterous hands where space constraints are less extreme, allowing the linkages to be designed with higher strength. However, for human-sized, high-DoF dexterous hands, the limited space makes it difficult to provide sufficient safety margins for linkages and hinges, leading to a higher risk of damage under frequent industrial use.
Another approach is direct drive, where miniaturized motors and reducers are placed inside the fingers, making full use of the finger and palm space. This allows for a good balance between DoF and overall size, but ensuring the reliability of small reducers during physical interaction and addressing the load and heat dissipation challenges of tiny motors pose significant difficulties.
Tendon-driven systems are the only one of the three transmission schemes that inherently offer compliance. We believe compliance is essential for the long-term stable operation of dexterous hands in industrial environments, as collisions and impacts with the physical world are almost inevitable. Given the size and material limitations of dexterous hands, purely rigid transmissions cannot achieve very high structural strength. To ensure sufficient robustness, the transmission system must have adequate compliance. However, traditional tendon-driven hands, like the Shadow Hand, often place the motors in the forearm to keep the hand as compact as possible. Our team has experimented with similar approaches, but this leads to an overly complex transmission system, prevents modular finger design, and creates significant challenges for mass production and reliability.
Ultimately, we prioritized practical considerations: ensuring system reliability, finger dimensions, and modular design, while accepting a slightly thicker palm. We believe this approach is more conducive to mass production and real-world use. Additionally, as motor technology advances, our next-generation products will be able to achieve a palm thickness closer to that of a human hand.
The “inner rigid, outer compliant” design increases tolerance for error during exploration; can you describe a real failure where that compliance prevented catastrophic damage but exposed a deeper system limitation?
So far, we haven’t identified a systemic limitation in the “inner rigid, outer compliant” design—after all, the human hand follows a similar principle. If there is a limitation, it might be that this compliance could become a constraint if we ever need to achieve “super-rigid” grasps that exceed human capability. But for now, we believe we still have much to learn from the human hand, both in hardware and software.
If tactile sensing is the true data gateway for embodied intelligence, what scenario has proven that vision-led policies still outperform touch-first strategies, and what strategic risk does that reveal in your roadmap?
We don’t think it’s necessary to debate whether vision or touch should take the lead—both are crucial ways for robots to perceive the world. For embodied intelligence, in scenarios that don’t require dexterous manipulation, vision-led strategies are generally simpler and more cost-effective, such as for navigation or simple grasping. However, we believe the core value of embodied intelligence lies in dexterous manipulation, and to excel at that, tactile sensing is indispensable. The strategic risk, if any, would only materialize if our fundamental assumption is wrong—that is, if dexterous manipulation is not, in fact, the core value of embodied intelligence.
Reinforcement learning imposes physical wear on hardware; at what point does the economics of real-world trial become unsustainable, and what irreversible engineering compromise have you had to make to extend hardware life?
If hardware breaks down or suffers significant performance degradation within just a few weeks of reinforcement learning training, it would severely limit algorithm iteration efficiency and become prohibitively expensive in terms of both time and economics. To extend hardware life, we opted for a primarily tendon-driven design. This might seem counterintuitive, as many assume tendon systems have inherent durability issues. However, a well-designed tendon-driven system offers the best potential for achieving long life and high reliability in dexterous hands. A prime example is the Shadow DEX-EE, a three-finger, fully tendon-driven hand capable of withstanding prolonged RL training, though it sacrificed a human-like five-finger configuration for durability. Our Apex Hand has already passed millions of load-bearing fatigue cycles without showing significant hardware wear, demonstrating the stability and reliability of our technical approach.
In semi-structured factory deployments, what deployment has underperformed expectations despite meeting lab benchmarks, and what did that reveal about the hidden assumptions in your product positioning?
Our product has only recently begun batch delivery and hasn’t yet been deployed at scale in factory environments, so we haven’t encountered a case where it met lab benchmarks but underperformed in the field. We anticipate the most significant gap could come from “environmental cleanliness.” Lab settings are relatively clean, whereas factories may have oil, grease, or dust—for instance, in automotive parts plants. These contaminants could affect the sensitivity of tactile sensors, demanding higher levels of sensor protection. This is a key area of focus for our next-generation product optimization.
As a platform provider supplying hardware, software, and baseline data, how do you navigate the tension between enabling customer differentiation and preserving your own long-term moat, and what moment forced you to confront that boundary?
The three core challenges for dexterous hand adoption are hardware, data, and models. Our positioning is as a professional dexterous hand solutions provider, focusing on solving the hardware and data collection tooling problems. Data and models are highly scenario-dependent and should be addressed by customers according to their specific needs—this is also where differentiation between embodied intelligence customers will emerge in the future. As model capabilities advance, they will demand hardware capable of more complex tasks, which in turn drives the need for next-generation hardware (e.g., with more sensitive touch). Our moat lies in our ability to continuously deliver products with the best performance, highest reliability, and most competitive cost. On a deeper level, it depends on our company’s iteration speed and execution capability. Only by excelling in these areas can we maintain our central position in this virtuous cycle.
If a vertically integrated humanoid manufacturer scales faster than expected and internalizes its own hand design, where does that leave Dexcel strategically, and what risk are you consciously accepting by remaining modular?
This scenario is certainly possible. However, we firmly believe in “specialists doing specialized work.” The core focus for humanoid robot manufacturers is whole-machine integration and scenario deployment, while the dexterous hand is an extremely precise subsystem. Developing a dexterous hand in-house would distract from their primary focus and slow their overall progress. Our value proposition is to help them validate their business models faster by providing a mature, ready-to-use dexterous hand product.
Hardware scale depends on supply chain stability and field maintenance; what single operational bottleneck would most likely constrain you in a sudden order surge, and when did you first recognize that vulnerability?
The most critical bottlenecks are in the supply chains for miniature reducers and flexible tactile sensors. Controlling the yield rates and ramping up production capacity for these core components are the key constraints on our delivery speed. We recognized this vulnerability as soon as we entered the mass production phase. Therefore, we approach the mass production and delivery of dexterous hands with utmost respect—it is by no means an easy task.
Many companies claim 2026 will be the commercialization inflection point for humanoids; what non-technical factor could derail that timeline for you specifically, and what contingency have you quietly prepared for?
Regulations and social acceptance are the key variables. Widespread public skepticism regarding the safety and ethics of robots could significantly delay commercialization. We are actively participating in the development of industry standards to help build a trustworthy image for robotics.
If tactile intelligence truly becomes foundational to embodied systems, what structural shift would need to occur across the broader robotics ecosystem for that thesis to prove correct, and what would failure of that shift look like five years from now?
If tactile intelligence becomes foundational, it would mean robots evolve from passively adapting to their environment to actively shaping it—transitioning from performing fixed tasks in structured settings to mastering delicate operations in unstructured environments. For this thesis to hold, the broader robotics ecosystem must undergo structural upgrades in several areas: mass production of tactile sensors, development of manipulation algorithms that effectively fuse tactile and other sensory modalities, and the emergence of viable commercial application scenarios. If this shift fails, in five years, robots will likely still be unable to enter domains requiring fine manipulation or home environments. The core value of embodied intelligence—dexterous operation—would remain unrealized, trapped in the lab. The total addressable market would shrink drastically, investment and talent interest would wane, ultimately plunging the industry into a prolonged winter.
Editor’s Note
This interview examines the shift from mobility to manipulation in embodied robotics, where tactile sensing and hardware reliability define practical deployment.
