A functional dexterous hand is no easier to build than a robot itself. It endures every hardship a robot faces, yet must achieve perfection at an even smaller scale.

As a core component of humanoid robots, a dexterous hand refers to an end-effector designed to mimic the kinematics of human hands. Compared to grippers, dexterous hands offer greater versatility, enabling fine operations such as screwing and grasping irregular objects. By using dexterous hands, robots can utilize various tools found in human daily life.

However, according to 2025 statistics, there is a significant disparity between global humanoid robot body manufacturers and dexterous hand manufacturers in terms of quantity and valuation. Data from the New Strategy Humanoid Robot Industry Research Institute indicates that as of April this year, the number of global humanoid robot body enterprises has exceeded 300. M2 Mitu Consulting data shows that as of July this year, there are already 122 global dexterous hand enterprises.

In terms of valuation, Figure, a humanoid robot enterprise that previously completed Series C funding, has reached a post-money valuation of $39 billion (approximately RMB 270 billion). After Unitree Robotics completed its Series C funding in June, its valuation exceeded RMB 12 billion. Meanwhile, leading dexterous hand manufacturer Shadow Robots has a valuation exceeding RMB 2 billion, with other companies' valuations mostly around RMB 1.5 billion.

Currently, although the number and valuation of dexterous hand enterprises are relatively low compared to those of humanoid robots, their role as critical hardware on robots is immeasurable.

Professor Sun Fuchun from the Department of Computer Science at Tsinghua University has pointed out: "Robots represent the 'last mile' of automation, while dexterous hands are the 'last centimeter' of robots."

However, truly building a stable and functional dexterous hand is a difficult task. The 'impossible triangle' of cost, performance, and reliability has long plagued the entire industry.

In this regard, Elon Musk is also deeply troubled; he has revealed that the engineering effort for Optimus's dexterous hand may account for half of the total robot development. At a previous ALL-IN summit, Musk stated, "If you want robots to truly be able to perform all tasks humans can do, you must first solve the problem of the dexterous hand."

Specifically, achieving the necessary degrees of freedom and precision in a dexterous hand requires significant R&D efforts in actuation, sensing, and transmission structures. Integrating motors, transmissions, and sensors within the limited space of the hand makes it extremely difficult to achieve a perfect balance between hand weight, heat dissipation, and power density.

Moreover, with each additional degree of freedom added to the hand, the complexity of control and reliability challenges increase exponentially. Consequently, current dexterous hands struggle to simultaneously meet the requirements for high performance, reliability, and low cost.

However, as a key hardware component for general-purpose humanoid robots, the dexterous hand remains one of the most valuable parts of the body, typically accounting for about 20% of the total unit value.

This means that the dexterous hand is a critical hardware component that significantly impacts the overall cost structure and commercialization timeline. Its technological maturity and supply chain readiness are major bottlenecks determining whether humanoid robots can transition from being merely functional to truly useful and practical.

Furthermore, the dexterous hand sector will benefit substantially as the pace of large-scale humanoid robot mass production changes. The GGII (High-Work Robot Industry Research Institute) previously predicted that by 2030, global sales of humanoid robots would approach 340,000 units, with a market size exceeding 64 billion yuan; by 2035, global sales would exceed 5 million units, with a market size surpassing 400 billion yuan.

If more than half of all humanoid robots opt to equip dexterous hands, the market potential becomes evident. Under this vast future market, competition among current dexterous hand manufacturers hinges on technological breakthroughs: those who best balance product cost, performance, and reliability will have the opportunity to capture a larger share of the cake. However, honestly, this is not as simple as it seems; creating a truly effective dexterous hand remains a long and arduous journey.

The technical path has not yet been finalized

In fact, dexterous hands and robots are not products that have only emerged in recent years. It is the rapid development of embodied AI technology that has allowed them to come more into the public eye.

As early as the 1970s, the Japanese Electrical Laboratory introduced the Okada three-finger dexterous hand, which featured a joint layout and tendon-driven system similar to that of the human hand while possessing 11 degrees of freedom.

Later, in the 1980s, Stanford's JPL three-fingered nine-degree-of-freedom dexterous hand, which employed twelve DC servo motors and a tendon-driven system, laid a certain foundation for the transmission systems of modern dexterous hands.

However, early dexterous hands had relatively low degrees of freedom and were predominantly three-fingered, with even higher costs, lacking the foundation for mass production.

By the 2000s, Shadow Hand from the UK's Shadow Robot increased its degrees of freedom to 24. The technology for dexterous five-fingered hands also became more mature, enabling complex operations such as unscrewing bottle caps and achieving small-batch mass production.

In the 2010s, the intelligent capabilities of dexterous hands saw further enhancements. A representative example is OpenAI's Dactyl, which can autonomously solve a Rubik's Cube through reinforcement learning. From this point onward, dexterous hands began to take on the雏形 (embryonic form) of modern products that are both flexible and intelligent.

By 2022–2023, Tesla’s Optimus dexterous hand had iterated from 11 degrees of freedom (DOF) to 22 DOF. It adopted a structure combining planetary gears, ball screws, and tendon cables, enabling complex actions such as tossing and catching a tennis ball.

In the subsequent years, with technological advancements, the technology behind dexterous hands gradually matured, leading to several relatively stable solutions. The synergy among actuation, transmission, sensing, and control remains key to achieving high precision and perception in dexterous hands.

As the power source for dexterous hands, current actuation methods include motor-driven, hydraulic-driven, and pneumatic-driven systems. Currently, most dexterous hands primarily use motor-driven solutions, including coreless motors. Transmission refers to the method of transmitting power, with options such as tendon-driven, linkage-driven, and worm gear drives. Sensing and control mainly involve the decision-making and execution steps.

Among these components, transmission offers the most diverse technical routes within the industry. Tendon-driven systems offer advantages such as lightweight design, spatial flexibility, and ease of achieving multiple degrees of freedom. However, their major drawback is susceptibility to wear and tear, along with complex maintenance requirements.

Linkage-driven systems boast high rigidity, large load-bearing capacity, and stable transmission efficiency, though they are highly sensitive to space constraints and assembly precision. Worm gear drives provide fast response times and high positioning accuracy, but they tend to be bulky and heavy.

But the industry is also exploring compound drive systems, which use arrangements of tendon cables, gears, and other components to select relatively superior solutions. This approach can combine the advantages of various schemes, although it significantly increases complexity and integration difficulty.

Overall, there is currently no standard answer for transmission solutions; choices are instead matched to factors such as task load and cost.

In specific products, Lingxin Qiaoshou's Linker Hand L20 adopts a linkage transmission method with self-developed motor drives. It can simulate a natural grasping motion similar to that of human hands, enabling richer and more precise operations.

Boya Intelligent's Gaoshan D22Pro dexterous hand, on the other hand, uses a tendon cable transmission scheme. The advantage of this solution is lighter fingertip weight and faster response speed.

Additionally, regarding worm gear technology, Taoshi Intelligent's toroidal enveloping reducer can achieve an overall hand load capacity of 50 kg and a motion accuracy of ±0.015 mm in robot dexterous hand applications.

Reliability Is Key to Implementation

Elon Musk has publicly discussed the challenge of dexterous robot hands on multiple occasions. During Tesla's Q3 earnings call, he noted that creating a hand as flexible and capable as a human hand is not easy, stating that Optimus's hand and forearm represent a significant engineering challenge. Previously, LatePost reported citing relevant sources that "the service life of Optimus's dexterous hand does not exceed two months; the flexible electronic skin covering its fingers and palm accelerates wear when Optimus touches objects, and the tendon cables driving the fingers are prone to aging and breakage."

This highlights the reliability issue within the 'impossible trinity' of dexterous hands: they are not as durable in actual use as imagined and can even overheat during operation.

Musk's anxiety is not isolated; this is a headache for the entire industry.

For dexterous hands to work effectively on robots, the first hurdle is size. Typically, they are designed to mimic the dimensions of a human hand. If the size is too large, it reduces collaborative efficiency with the robot body. Furthermore, larger dimensions mean greater weight, which negatively impacts the robot's effective payload capacity.

A contradictory problem arises because high-degree-of-freedom dexterous hands are difficult to miniaturize. Integrating hundreds of micro-motors and other components requires them to work together without interference, demanding extremely high levels of component miniaturization and integration.

During this highly integrated process, internal space is squeezed, forcing motors, sensors, and other parts to be packed tightly together, inevitably leading to concentrated heat generation. However, there is insufficient space inside the hand for cooling systems, making it difficult to create small dexterous hands equipped with adequate thermal management.

If efforts are made on thermal design, operations such as increasing the thickness of the casing or adding heat dissipation fins will in turn increase the volume of the dexterous hand. This would affect performance metrics such as agility, making it extremely difficult to strike a balance.

Moreover, the heating issue of dexterous hands typically arises during high-frequency, high-precision work, which generally occurs in factory settings where robots are intended to densely replace human labor. Heat causes a decline in the hand's performance and, in extreme cases, can even lead to shutdowns. Repetitive tasks in industrial scenarios accelerate the wear and aging of dexterous hands. If a dexterous hand lasts only two months and faces risks of shutdown during operation, its reliability is clearly insufficient for industrial scenarios aiming to replace human labor.

Dexterous Hands Also Lack Data

Dexterous hands face data shortages just like robot bodies do, a problem that many other components do not have to confront. Therefore, the difficulty in manufacturing dexterous hands is no mere talk.

Although robots' understanding capabilities in areas such as vision and decision-making are advancing rapidly, it remains challenging for them to achieve human-like flexibility in the real world, especially regarding dexterous hand operations. Tasks that are very simple for humans, such as folding clothes or holding eggs steadily, appear exceptionally difficult for dexterous hands.

This is because dexterous hands lack large amounts of interaction data, particularly crucial force control data and data involving contact with various materials.

However, current data collection for dexterous hands faces a triple dilemma of scale, authenticity, and generalization. Mainstream collection methods often struggle to balance all three. Teleoperation-based collection schemes can obtain high-precision, high-authenticity data, but problems of low efficiency and high costs persist.

While learning from publicly available data such as videos is cost-effective and easy to scale, it lacks core information like tactile feedback and proprioception, resulting in limited data value.

Additionally, while simulation can rapidly generate massive amounts of data, the physical discrepancies between simulation and reality make direct application difficult, leading to insufficient generalization capabilities.

In real-world physical scenarios, dexterous hands face more complex manipulation objects. Since many physical shapes in daily life are irregular with varying forms, materials, and weights, laboratory settings cannot cover all possible scenarios. However, in more vertical or specialized scenarios, such as factories, where robotic operations are relatively fixed, the data challenges for dexterous hands are less significant, and the primary issue becomes hardware.

But the original intent of designing dexterous hands was to enable versatility, given that most real-world infrastructure is designed for human morphology. With dexterous hands, robots gain a physical foundation applicable across any scenario, making the resolution of universal data requirements essential.

To address this, the industry is gradually shifting toward a hybrid model combining real-world scene data collection with simulated data augmentation. This approach ensures physical authenticity while leveraging simulation to expand data volume and enhance generalization.

Meanwhile, algorithmic innovations enable efficient data filtering and compression, extracting core valid information to reduce storage and processing costs.

Previously, DexCraft open-sourced its large-scale multimodal dexterous manipulation dataset, DexCanvas, in October. This initiative introduces the critical dimension of haptic feedback into large-scale datasets, enabling robots to not only mimic human motion trajectories but also understand the mechanics involved in manipulation.

This represents their response, to some extent, to the challenge that existing datasets struggle to simultaneously achieve scale, realism, and haptic information.

Breakthrough Exploration Has Not Stopped

The current situation is that while everyone recognizes dexterous hands as a key component for enhancing robots' general capabilities, no one dares to claim they have truly created a product that balances cost, performance, and reliability. This is precisely why we still see many robots using grippers in various scenarios today.

However, despite the difficulties, dexterous hands remain an indispensable part of robotic components. For robots to achieve mass production, they must ultimately overcome this core hardware bottleneck.

Therefore, the entire industry is actively seeking solutions to balance the 'impossible trinity' of dexterous hands, aiming to make them truly practical and reliable products.

Technically, the industry is increasingly adopting biomimetic design combined with lightweight integrated structural solutions. By drawing on human hand anatomical features to optimize joint distribution and transmission structures, these designs maintain multiple degrees of freedom while reducing overall weight through lightweight materials such as carbon fiber, thereby addressing the issue of dexterous hands being too heavy.

Furthermore, hybrid transmission schemes that combine distributed motor drives with tendon-driven technologies have rapidly entered the stage. These approaches preserve the precision of independent joint control while enhancing operational compliance through flexible transmission, thereby reducing the risk of collision impact.

Additionally, integrating tactile, force, and position sensors at the fingertips and finger pads to improve operational reliability has become increasingly common. This strategy of multimodal perception fusion compensates for information deficits inherent in single-modal robotic systems and is key to enhancing the operational reliability of dexterous hands.

These continuous iterations at the technical level are essentially proactive breakthroughs by the industry paving the way for mass production. After all, technological advancements in dexterous hands significantly influence the deployment efficiency and market competitiveness of robots.

In short, regardless of the specific path taken, with many robot manufacturers having outlined future mass production plans, the dexterous hand industry stands at a crossroads in the era of robotics. Yet, in the pursuit of truly effective dexterous hands, no one chooses to give up.