As humanoid robots already walk freely in factories, the dexterous hands still being refined in laboratories have become the weakest link in the chain of embodied intelligence. As the "last centimeter" of interaction between humanoid robots and the physical world, dexterous hands are becoming the core technological high ground in the embodied intelligence track. Entering 2026, the dexterous hand market and capital have delivered a report card of nearly CNY 5 billion in Q1 financing, along with an optimistic forecast that annual sales will exceed 70,000 units. However, yet headwinds persist beneath this market boom—Tesla's Optimus has been repeatedly delayed due to hand bottlenecks, making the tough battle for mass production of dexterous hands much more difficult than imagined. Amid the intertwining of "ice and fire," the industry is turning its attention to the front lines of the dexterous hand industry: tendon-driven, linkage-driven, and direct-drive—which route can be the first to succeed? Cost, lifespan, and algorithms—which barrier is the hardest to overcome? As the industry stands at the critical point of transitioning from "making it" to "selling it," the answers to these questions are determining just how far embodied intelligence can go.
Technology Route Debate: No Single Standard Solution for Dexterous Hands
The technological routes for dexterous hands are far from converging at present, with the three major solutions—tendon-driven, linkage-driven, and direct-drive—each having their own followers. Among the companies betting on tendon-driven solutions, their starting points and implementation paths vary significantly. Zhou Chen, Co-founder and CEO of Dexterous Intelligence, bases his logic on "software-defined hardware"—"Embodied intelligence is different from the logic of automobiles or industrial robots, which prioritize hardware before software. It is software-defined hardware, or more accurately, model-defined hardware." Tendon-driven mechanisms are naturally closer to the driving method of the human hand, offering better responsiveness to the uncertainties of the physical world. At the same time, tendon-driven hand solutions can achieve higher power density per finger, greater cost reduction potential for the entire hand, and lower energy consumption when integrated with robots. Interstellar Lightyear focuses on tendon-driven dexterous hand technical solutions and is more optimistic about the "performance ceiling" of tendon-driven mechanisms. Its Co-founder Zhou Jiaqi stated bluntly, "Tendon-driven technology is indeed several orders of magnitude more difficult than direct-drive and linkage-driven solutions, and the requirements for control algorithms also increase exponentially, but its performance ceiling is much higher.
Tesla's announcement of a full shift to the tendon-driven route in the middle of last year actually reflects its recognition of this physical ceiling." However, Interstellar Lightyear is also clearly aware that a single drive route cannot cover the market demand for all scenarios and dimensions. Therefore, focusing on scenario pain points, it considers tendon-driven to be the optimal route more suitable for industrial implementation at this stage. It has made technological breakthroughs in algorithms and hardware structures on the tendon-driven route, such as the "drive-unit and end-effector separation" solution and the "fully automatic tendon self-tensioning algorithm," while also deploying a direct-drive product line as a supplement. Pacini Sensing has chosen a completely different linkage-driven solution. In the view of Lin Tong, Director of the Research & Integration Center at Pacini Sensing Technology (Tianjin) Co., Ltd., the tendon-driven solution has strong explosive power but a limited lifespan, while the direct-drive solution has excellent control but an insufficient output ceiling. In comparison, the linkage-driven solution is the optimal choice that balances high performance, long lifespan, and high stability, better fitting the commercialization needs of industrial implementation and long-term deployment at this stage.
He emphasized that the convergence of technological routes depends not solely on the mechanical structure, but on whether the algorithm can truly master the hardware. The ultimate competition for dexterous hands is not just about single physical parameters, but the comprehensive performance of the entire software and hardware system in real, complex scenarios. BrainCo has chosen the unique route of "full direct-drive + back-drivable." He Xiyujin, Partner at BrainCo, stated that this choice is not merely a hardware-level decision, but a result deduced from the perspective of algorithm trainability. Back-drivable joints allow synchronous collection of motion trajectory and contact force data during operation, naturally adapting to the training needs of imitation learning and reinforcement learning, and significantly reducing the difficulty of migrating from simulation environments to real-world deployment. "The future development of embodied intelligence increasingly relies on imitation learning and reinforcement learning, and hardware capabilities directly determine the ceiling for robots to learn and execute actions," she said. Reviewing the technological route choices of multiple enterprises, they essentially correspond to three different industrial philosophies: tendon-driven pursues the ultimate physical performance ceiling, linkage-driven bets on engineering reliability at this stage, and direct-drive yields some performance space for the trainability of algorithms. Experts point out that in the next 2-3 years, the industry will not quickly converge to a single route, but will form a differentiated development pattern based on application scenarios. This means that for a considerable period, the market will support multiple technical solutions in parallel, and the evolution and iteration of each will be more meaningful than "who is right and who is wrong."
However, no matter which route is chosen, the industry must face a more cruel reality: if it cannot be made, cannot be made cheaply, and cannot last long, even the most perfect technological solution can only remain in the laboratory. The repeated delays of Tesla's Optimus precisely confirm that between technological verification and scaled mass production, there lies a gap far more perilous than route selection.
From Prototype to Commercialization: The Mass Production Dilemma
As dexterous hands move from the laboratory to mass production, they face challenges at three levels: technological R&D, engineering mass production, and market acceptance. First is the breakthrough of technological limits—miniaturization of drives, reliability of transmission, and coordination of perception and control algorithms, each of which is enough to constitute a "choke point" problem. Taking Tesla's difficulties in dexterous hand R&D as an example, Zhou Chen believes that the underlying logic for Tesla's delay in mass production and open-sourcing of dexterous hand designs is the same: it underestimated the difficulty of mass-producing tendon-driven dexterous hands during the design phase. It still followed the traditional approach of separating the topics of functional realization and mass production cost reduction. However, for tendon-driven dexterous hands, mass production feasibility must be taken as an important design constraint from the very beginning. Zhou Chen further pointed out that the "hardest bone to chew" is actually the mapping and adaptation between perception hardware (tactile sensors) and drive models. He stated that it is difficult to approximate human touch and dexterous hand touch in the short term. The most difficult part is how to transfer and reproduce the closed loop from perception to manipulation of the objective physical world by the human hand onto the different sensors and manipulation capabilities of the dexterous hand. "Just as pigeons can navigate by feeling the Earth's magnetic field with their bodies, humans need to rely on a compass to achieve the same navigation purpose. The two entities are different, but the goals are the same. This kind of transfer is the most difficult part." He Xiyujin echoed this judgment from the perspective of systems engineering: "These problems do not exist in isolation in essence.
If one must say what the 'hardest bone to chew' for dexterous hands is, I think it is achieving comprehensive capabilities close to the level of the human hand within a limited volume." Robot dexterous hands need to integrate drive, transmission, perception, and control systems simultaneously within a very small space. She pointed out that for embodied intelligence, the real challenge lies in enabling robots to not only "move" but also "perceive" and "learn." Therefore, the key to future dexterous hand competition will not be the leadership in single component parameters, but the closed-loop integration capability of mechanical structures, sensing systems, and intelligent algorithms. While technical challenges define the ceiling of R&D capability, mass production validates the maturity of engineering systems. Multiple interviewees unanimously regard "batch consistency" as the core bottleneck. Lin Tong stated bluntly that dexterous hands are essentially high-density precision electromechanical systems.
How to ensure the long-term stability, reliability, and cross-batch consistency of each hand in massive production is the real key to determining the speed of industrial implementation. He Xiyujin further pointed out that it is not difficult to make a dexterous hand with excellent performance in the laboratory. The difficulty lies in still being able to ensure performance consistency, reliability, and controllable costs when producing tens of thousands of products. Behind this is the test of comprehensive capabilities in design, process, supply chain, and quality management. "In the future, the focus of industry competition will gradually shift from single-point technological innovation to the construction of engineering and industrialization capabilities." Going a step further, even if technological breakthroughs and engineering systems are in place, there is a more fundamental question that needs to be answered—why does the market need dexterous hands? The current dexterous hand industry is similar to the early development stage of smartphones. Given the gradual improvement of the supply chain and the popularization of platform-based solutions, the industry will usher in a clear inflection point for cost reduction. Lin Tong believes that in the next 3-5 years, the overall cost curve of dexterous hands will enter a steady downward range. When dexterous hands with high degrees of freedom, fine force control, and high mobility capabilities have their single-hand cost gradually controlled between CNY 5,000 and CNY 10,000, the industry will usher in true large-scale popularization and explosion. Notably, multiple interviewees emphasized a counter-intuitive judgment: cost is not the biggest obstacle to popularization. Zhou Chen's metaphor is the most sharp: "Cost is not the main barrier to dexterous hand adoption. Even if dexterous hands were given away for free today, widespread deployment would remain impossible. Just like if there are no drivers, giving away cars for free won't get anyone on the road." He believes that the key from "making it" to "selling it" is "the capability of integrated software and hardware dexterous manipulation models, and the challenges mainly come from data and models."
He Xiyujin added an easily overlooked dimension—the hidden costs for users. The actual cost for customers is not just the procurement price itself, but also the development time and manpower investment required from receiving the product to truly running through the application scenarios. For many small and medium-sized enterprises and research institutions, this part of the hidden cost is often no less than the hardware itself. In summary, the mass production dilemma of dexterous hands is not a single issue of "technological breakthrough" or "cost reduction and efficiency increase," but a systematic project involving technological breakthroughs, engineering system construction, and user ecosystem cultivation. The absence of any link may become a bottleneck for industrialization.
Beyond Spec Wars: Redefining "Dexterity" Through System-Level Performance
As mass production issues are gradually resolved, a deeper question emerges: what kind of hand can be considered "truly dexterous"? The industry is undergoing a conceptual shift from "parameter worship" to "task capability." Currently, the degrees of freedom (DOF) of dexterous hands on the market range from 6 to 42, but the interviewed enterprises unanimously agree that the piling up of DOF does not equate to an improvement in dexterity. "Strictly speaking, 6 DOF does not count as a dexterous hand. 42 DOF is actually redundant design." Zhou Chen analyzed from a biomechanical perspective that the 22 DOF of the human hand is essentially to achieve the locking and adjustment of the 6 DOF of any object in the 3D world, requiring at least three three-joint fingers and at least 12 DOF. The truly necessary DOF is around 20. Lin Tong pointed out from a control perspective that DOF does not directly equate to dexterity. When the DOF exceeds a certain threshold, blindly piling up DOF will not only lead to increased hardware costs but also significantly increase the complexity of control algorithms, which may instead reduce the success rate of tasks in actual operations. He Xiyujin also emphasized that there is a common misconception that more DOF means more advanced. In fact, if the control capability cannot keep up, excessive DOF will instead increase system complexity and failure rates. If the discussion on DOF reflects the industry's reflection on "hardware indicators," then the coordination between dexterous hands and the robot's "brain" further pulls the discussion from hardware to the system level. Zhou Chen pointed out that the coordination between dexterous hands and the robot's brain requires deploying "perception/manipulation" models on the edge side, so that the specific actions in the task chain decomposed by the brain can truly be completed in the physical world. It is reported that Dexterous Intelligence has focused on data collection of human hand perception and manipulation, as well as edge-side model training paradigms, and the results have been open-sourced.
He Xiyujin emphasized that no matter how advanced a dexterous hand is, without intelligent decision-making capabilities, it can only complete preset actions; and no matter how powerful a large model is, if it cannot interact with the physical world through dexterous hands, it is difficult to truly achieve embodied intelligence—the key to future competition is the closed-loop capability from perception and decision-making to execution. "One of the most serious misconceptions in the industry in the past is blindly treating dexterous hands as an appendage to humanoid robots, thereby severely underestimating its independent market value." Zhou Jiaqi stated that the market positioning of dexterous hands can not only be a part of humanoid robots but is more likely to achieve a commercial closed loop in industrial scenarios first as an independent "manipulation terminal." As the path of software and hardware coordination gradually becomes clear, the industry will inevitably face the next question: how to enable hands from different manufacturers, robots from different brands, and different algorithm models to be interconnected? This involves standardization. Zhou Chen believes that the hardware interfaces and communication protocols between dexterous hands and robotic arms are easy to standardize. But this kind of standardization is superficial; the more underlying layer should be the standardization of perception, data, and models. And it is too early for software standardization. Just like AI large models, there are currently no global standards. Technology cannot be standardized during the period of rapid development and trial and error. Only after the technology meets market needs is standardization required to reduce usage and replacement costs. He Xiyujin proposed a pragmatic path: first form de facto interoperability at the software and data levels, and then gradually extend to hardware interfaces. She pointed out that the ultimate goal of standardization is not to unify product forms, but to reduce collaboration costs and improve industrial efficiency. In this process, interoperability at the software level is often easier to achieve than the unification of hardware interfaces. The above judgments provide a more flexible standardization idea for the industry, which is still in the technological evolution period—not rushing to unify hardware forms, but first lowering the collaboration threshold at the development and data levels. Following this line of thought, the judgment on the end game of the industry is also gradually becoming clear: there are currently three approaches in the dexterous hand track—independent integrators, in-house R&D by complete machine manufacturers, and downward extension by component enterprises. Zhou Chen judges that in the future, products from independent integrators will be the mainstream choice for OEMs.
A few complete machine manufacturers will have the ability to develop their own dexterous hands, and component enterprises will return to their professional component production capabilities. He Xiyujin predicts that the industry will show a "polarization" trend: top robot complete machine manufacturers will insist on in-house R&D of core components to gain system-level advantages; on the other hand, a group of specialized dexterous hand enterprises will also emerge, becoming industry-level suppliers and providing solutions for multiple robot platforms. Although the two judgments differ, they point to a common trend: independent dexterous hand enterprises will occupy an important position in the industrial ecosystem. Their core competitiveness lies not in a single technological breakthrough, but in whether they can simultaneously possess core technology, mass production capabilities, and ecological cooperation capabilities. Regarding the breakthrough direction in the next 3-5 years, the judgments of multiple interviewed enterprises are quite consistent—tactile perception.
Lin Tong believes that "the popularization of high-precision touch" will be the most decisive breakthrough. In the real physical world, the vast majority of fine manipulations essentially cannot be separated from the closed-loop feedback of touch. The core gap between future humanoid robots lies in the deep understanding of the interaction logic of the physical world, and high-precision tactile perception is the key to unlocking this core capability. As for the true "volume explosion moment" of dexterous hands, all enterprises have maintained a prudent attitude. Zhou Chen stated that the volume explosion moment for dexterous hands will be after the convergence of dexterous manipulation model technology; hardware maturity is a necessary but not sufficient condition. Lin Tong believes that volume explosion requires two conditions to mature simultaneously: first, self-iteration on the supply side, where dexterous hands must meet strict industrial-grade standards in terms of working lifespan, production costs, and operational reliability; second, the scale effect on the demand side, where the delivery of tens of thousands of units in the humanoid robot market forms a powerful scale traction. He Xiyujin gave a relatively optimistic judgment—the true volume explosion moment requires three conditions to mature simultaneously: first, clear application scenarios where robots can continuously create commercial value; second, a decrease in complete machine costs to reach market-acceptable levels; third, sufficiently mature dexterous manipulation capabilities to stably complete complex tasks in real environments. If these three conditions are gradually met, it is expected that the dexterous hand market will enter a stage of rapid growth in the next 3-5 years.
Author | Yang Pengyue Editor | Wu Lilin Art Editor | Maliya Supervisor | Lian Xiaodong