1X Unveils NEO Robotic Hand: 25 DoF Tendon-Driven Design Ushers Dexterous Hands into the Mass-Production Era

The "Hand" Is Becoming the New Battleground for Humanoid Robots

Competition in humanoid robotics is shifting from overall form factor toward refined components, and the "hand" is without doubt one of the hardest to master—and most commercially valuable—elements. Recently, robotics company 1X unveiled a brand-new dexterous hand for its NEO humanoid robot. With 25 degrees of freedom, a tendon-driven architecture, and full-joint force control among a raft of design features, it sets a new benchmark for dexterous hand engineering. According to information compiled by @TheHumanoidHub, this robotic hand is not only aggressive in its specs but also holds mass-production potential on the manufacturing side.

1X Technologies' Strategic Positioning: 1X Technologies (formerly Halodi Robotics) is a Norwegian-American robotics company that has received strategic investment from institutions including OpenAI. Unlike many humanoid robot companies that focus on industrial scenarios, 1X has targeted home services as its core market from the very beginning, and the NEO series is the product embodiment of this strategy. This positioning directly shapes the design orientation of the dexterous hand: home scenarios require a robot capable of handling everything from eggs to hammers, while guaranteeing absolute safety in environments shared with humans. In the dexterous hand space, current major competitors include companies focused on end effectors such as Sanctuary AI and Dexterous Robotics, as well as in-house hands developed by full-system manufacturers like Tesla Optimus, Figure 02, and Unitree G1. Most mass-produced humanoid robots today still use simplified grippers with 3–5 degrees of freedom, and the 25-DoF solution represents a major leap in technological approach—but it also means an order-of-magnitude increase in control complexity and cost.

For humanoid robots, bipedal walking has long ceased to be the biggest bottleneck. What truly determines whether they can enter homes and factories to accomplish real tasks is a pair of hands capable of grasping, manipulating, and sensing. 1X's release is precisely an answer to this core question.

25 Degrees of Freedom: Biomimicry Over Uniform Distribution

The new robotic hand has a total of 25 degrees of freedom, of which 22 are distributed across the fingers and palm, and the remaining 3 are in the wrist. These degrees of freedom are not evenly distributed—they follow a deliberately "biased" design based on human anatomical structure.

The Engineering Meaning of Degrees of Freedom: Degrees of Freedom (DoF) is a core concept describing the independently movable dimensions of a mechanical system. The human hand has roughly 27 degrees of freedom, spanning flexion-extension, adduction-abduction, rotation, and other modes of motion. In engineering terms, each additional degree of freedom means an extra actuation unit, sensor, and control logic, with complexity growing nonlinearly. Twenty-five degrees of freedom is a top-tier configuration among commercial robotic hands—Boston Dynamics' early Atlas hands had only a handful of DoF, and while Shadow Robot's biomimetic hand reached 24 DoF, it cost on the order of a hundred thousand dollars and was difficult to mass-produce. The value of degrees of freedom lies not in the number itself, but in whether their distribution matches actual manipulation needs.

Thumb Opposition Capability Is Specially Enhanced

1X deliberately tilts resources toward the thumb, enabling it to form true opposition with the other fingers. This may seem like a detail, but it is in fact crucial—the core of human hand dexterity comes precisely from the thumb's opposition capability, which makes fine grasping, pinching, and tool use possible. From the perspective of human evolution, thumb opposition is one of the core traits distinguishing the genus Homo from other primates; it is precisely this capability that enabled humans to use tools and build civilizations. In robotic hand design, many early approaches adopted uniform DoF distribution, thereby sacrificing the very functionality most central to actual manipulation. By choosing "biomimicry first," 1X signals that its goal is to have NEO accomplish complex real-world tasks rather than pursue symmetrical aesthetics in the lab.

The Engineering Significance of Grasp Taxonomy: Human grasping actions can be divided into two broad categories: "power grasp" (such as clenching a fist around a hammer) and "precision pinch" (such as pinching a needle and thread), and the two make entirely different demands on hand structure. Power grasp relies primarily on the overall envelopment of the palm and four fingers, while precision pinch depends heavily on the coordinated opposition of the thumb with the index and middle fingers. Research shows that more than 60% of humans' daily manipulation tasks require some form of precision pinch. This means that if a robotic hand's thumb opposition is insufficient, its coverage of daily tasks will be fundamentally limited—no matter how excellent its other specs are. 1X places thumb enhancement at the top of its priorities precisely because of a deep analysis of real-world usage scenarios, rather than mere engineering showmanship.

Tendon Drive and Low Gear Ratio: Light, Fast, and Powerful

On the transmission architecture front, this robotic hand uses a self-developed tendon-driven, quasi-direct-drive design, with a gear ratio of only about 5:1 to 15:1—far lower than the industry-common 100:1 to 200:1.

The Biomechanical Origins of Tendon Drive: The tendon-driven architecture directly mimics the biological structure of the human hand—there are no muscles in the human hand; the fingers are driven by flexor and extensor tendons extending from the forearm. The engineering advantage of this design is that it places the heavy actuation units (motors) far from the end effector, thereby dramatically reducing the finger's rotational inertia. The trade-off is that tendon elasticity and friction introduce nonlinearities, increasing the difficulty of precise force control. The application of tendon-driven architecture in robotics dates back to the 1980s, when Stanford University researchers first attempted to engineer this biomimetic principle; its core lies in achieving "decoupling of actuation and execution." However, this system faces three stubborn challenges: creep (tendons elongating after prolonged loading, causing precision loss), hysteresis (mechanical asymmetry between stretch and rebound), and friction accumulation. Early adopters of similar approaches included NASA's Robonaut 2, but these systems were extremely expensive and complex to maintain, and never reached mass production.

The Critical Role of Tendon Materials: Dyneema (ultra-high-molecular-weight polyethylene fiber, UHMWPE) is the mainstream material choice in the robotic tendon field today. Its specific strength is about 15 times that of steel wire of the same cross-section, its density is lower than water (it can float), and it offers excellent wear resistance and low creep characteristics—compared to traditional nylon or steel wire rope, its elongation under cyclic loading can be reduced by an order of magnitude. However, "low creep" is a relative concept; in high-frequency, high-load finger-drive scenarios, the accumulation of tiny amounts of creep will still cause systematic drift in tendon tension, requiring continuous compensation by control software. The core challenge in 1X's claimed mass production is precisely raising the manufacturing yield and service life of this precision system to a commercially acceptable level—this means it has made substantial breakthroughs in high-performance fiber material selection, tensioning-compensation algorithms, and assembly processes. If it can solidify these technical barriers, it will form a competitive moat that is difficult to replicate.

The Physical Principles of Quasi-Direct Drive and Backdrivability: The concept of quasi-direct-drive was popularized by MIT's Biomimetic Robotics Lab in the Mini Cheetah project, with theoretical roots tracing back further to Marc Raibert's team at Carnegie Mellon University in their 1980s–90s research on legged robots. Traditional industrial robots use high-gear-ratio (100:1 to 200:1) harmonic reducers or worm gears to output large torque from small motors, at the cost of a system that is completely non-backdrivable—when the gear ratio is 100:1, reverse transmission efficiency can drop below 10%, meaning external forces cannot effectively push the motor backward, which is a serious hidden danger in safety scenarios. The core of quasi-direct drive is to use a low gear ratio paired with high-torque-density motors, allowing the system to remain backdrivable while maintaining a certain torque output. The physical essence is: a low gear ratio means high reverse transmission efficiency, so external impact forces can be smoothly transmitted back to the motor end and be sensed or absorbed, rather than being abruptly blocked by the self-locking characteristics of a reducer. Migrating this technological paradigm from legged robots to dexterous hands is a classic case of cross-domain technology fusion.

The Engineering Dividends of Rear-Mounted Motors

The motors are placed in the forearm, driving the fingers by pulling tendons through the wrist. This design brings multiple advantages: the hand body itself is extremely lightweight with low end-inertia, while the larger-volume forearm motors ensure powerful output. Low inertia means finger movements are more nimble and responsive, while strength is supplemented by the rear-mounted motors.

More importantly, the low gear ratio gives the system natural "backdrivable" capability. Compared with traditional high-gear-ratio robotic hands—which tend to be rigid and hard to drive in reverse, prone to damage or injuring people when struck—1X's solution finds a better balance between dexterity and safety.

The Implicit Challenge of Motor Technology Selection: Placing high-torque-density motors in the forearm requires the motors themselves to achieve the utmost power density. Current high-performance motors in robotics generally use outer-rotor brushless DC motors (BLDC) paired with high-fill-factor windings and rare-earth permanent magnet materials (neodymium-iron-boron, NdFeB), achieving power densities 3–5 times that of traditional industrial motors. However, thermal management is a core challenge for high-power-density motors during continuous operation—forearm space is limited, heat dissipation paths are constrained, and overheating can lead to motor demagnetization and efficiency loss. If 1X is to integrate the motor array required for 25 degrees of freedom within the limited forearm space while ensuring thermal stability, it means it has made deep engineering investments in motor design, thermal interface materials, and heat-dissipation structures—which is precisely the core advantage of self-developed motors over off-the-shelf solutions.

Full-Joint Force Control and Tactile Sensing

Sensing capability is another major highlight of this robotic hand. All 25 degrees of freedom natively support force control and are fully backdrivable, with each joint simultaneously acting as a force sensor.

The Practical Value of Closed-Loop Proprioception

The company specifically emphasizes the importance of closed-loop proprioception: the robotic hand can "sense" its own posture and applied force at any time without relying on vision. Proprioception was originally a neuroscience term, referring to an organism's ability to perceive the position, motion state, and applied force of its own limbs, achieved through receptors such as muscle spindles and tendon organs. In robotics, this concept is borrowed to describe a system's ability to perceive its own state through internal sensors (joint encoders, torque sensors, etc.). Visual processing has latency (typically 30–100 milliseconds), while joint-level force control can respond to contact events on a millisecond timescale—this is precisely why humans can fumble to open a door in the dark, or type without looking at their hands. This is critical for robots operating in occluded, dark, or complex environments—it can complete tasks by "feel" just like a human.

The Technical Path to Force Control Implementation: There are two main engineering paths to implementing joint-level force control: one is to install six-axis force/torque sensors directly at the joints, which offers high accuracy but adds weight, volume, and cost, and the sensor itself is a potential failure point; the other is to indirectly estimate torque through current sensing (current-torque mapping), inferring joint force using the proportional relationship between motor current and output torque—low cost but with accuracy affected by factors such as friction and thermal drift. Because quasi-direct-drive systems have low gear ratios and high transmission efficiency, the accuracy of current-torque mapping is greatly improved—this is a key synergistic effect where the two technical approaches reinforce each other. 1X's phrasing that "each joint simultaneously acts as a force sensor" implies it primarily relies on the current-sensing path, achieving usable accuracy through precise calibration and algorithmic compensation—a choice that offers significant advantages in cost control and system simplicity.

The fingertips and contact surfaces are covered with tactile skin capable of simultaneously measuring contact force and shear force. Tactile skin is one of the long-unsolved commercial breakthroughs in robotics. Human skin contains roughly 2,500 tactile receptors per square centimeter, encompassing Merkel discs (fine touch pressure), Meissner corpuscles (light touch and texture), Pacinian corpuscles (vibration), and Ruffini endings (skin stretch), forming a multimodal, high-density sensing network. Current mainstream technical approaches each have limitations: piezoresistive sensors are mature in fabrication but insensitive to shear force; capacitive flexible sensors are highly sensitive but hard to calibrate; optical tactile sensors represented by GelSight offer extremely high accuracy but are limited in real-time performance; magnetic tactile sensors (such as Meta AI's ReSkin) have good durability but limited resolution. Shear-force detection is especially important because it is the key signal for judging whether "the grasp is about to fail"—before an object slips, tiny tangential displacement at the contact surface inevitably occurs, which is an order of magnitude faster than waiting for the object to actually slip and then detecting it visually. Top labs such as Stanford, CMU, and MIT have worked deeply in this direction for years, but mass-production challenges persist: poor durability of flexible materials, complex calibration, and high demands on signal-processing algorithms. The robotic hand can thus perform adaptive grasping in real time, dynamically adjusting grip strength based on object texture and slip state—neither crushing an egg nor dropping a tool.

The AI Training Value of Tactile Data: The commercial value of tactile sensors lies not only in real-time control but even more in data accumulation. When 1X's robotic hands complete millions of real-world manipulation tasks, the tactile time-series data from each grasp—contact-force buildup, shear-force changes, slip-warning signals—will form an extremely valuable training dataset of manipulation skills. Such data is currently extremely scarce in academia, because robotic hand systems equipped with high-quality tactile sensing have never been deployed at scale. If 1X can perform end-to-end reinforcement learning training combining tactile data with manipulation policies, it will form a positive flywheel of "hardware scale → data accumulation → model capability → hardware optimization"—this is the deeper logic of the "manufacturing as moat" strategy in the AI era. The combination of force control and tactile sensing constitutes the complete sensory closed loop for dexterous manipulation.

Safety and Durability: A Robot You Can "Wash Its Hands"

1X's investment in safety and durability is equally noteworthy. This robotic hand reaches IP68 waterproof rating and meets food safety standards—it can be rinsed directly, which is extremely practical in home cleaning and food processing scenarios.

The Industrial Meaning of the IP68 Protection Rating: The IP (Ingress Protection) rating is defined by the IEC 60529 standard. IP68 is its highest consumer-grade rating: the first digit "6" indicates complete dust protection, and the second digit "8" indicates the ability to withstand continuous submersion underwater (typically over 1 meter and over 30 minutes). For a robotic hand, achieving IP68 means the entire actuation, sensing, and electronic systems must undergo rigorous sealing design—which naturally conflicts with the open structure of tendon drive: tendons need to slide within conduits, while sealing requires that no water penetrate any moving interface. Resolving this contradiction requires meticulous engineering trade-offs in dynamic sealing materials (such as Teflon-coated conduits) and structural zone isolation.

The Market Signal of Food Safety Certification: Food safety standards (typically corresponding to regulatory frameworks such as FDA 21 CFR or EU EC 1935/2004) require that materials contacting food be non-toxic, must not migrate harmful substances, and must not have surface structures with hard-to-clean dead corners. This imposes additional constraints on the robotic hand's surface coatings, sealing materials, and structural design—for example, the lubricant in tendon conduits must be food-grade, silicone rubber seals must pass food-contact testing, and tactile skin materials must not contain migratory plasticizers. Combined with IP68 waterproofing, this design combination significantly expands the boundaries of application scenarios, implying that 1X is targeting housekeeping (kitchen cleaning, dishware handling) and food processing (food sorting, packaging) as early commercialization priorities—these two scenarios share the traits of high task repetitiveness, high labor-cost pressure, and moderate dexterity requirements, making them the ideal bridgehead for dexterous hands to move from the lab to commercialization.

Structural Compliance Delivers Intrinsic Safety

Thanks to the low gear ratio, tendon drive, and low end-inertia, the robotic hand possesses "compliant by construction" characteristics. When subjected to external impact, the fingers are safely driven in reverse rather than rigidly resisting the force. The company describes it as "actively yielding when struck by a hammer or caught in a drawer." This physical-layer safety guarantee is far more reliable than pure software force limits—software limits depend on sensors working properly and the control loop responding in real time, whereas structural compliance remains effective even in extreme situations such as system power loss or loss of control.

The Regulatory Background of Human-Robot Collaboration Safety Standards: Structural compliance is not only an engineering preference but also a prerequisite for regulatory compliance. ISO/TS 15066 (technical specification for collaborative robots) and the upcoming updated ISO 10218 series standards set explicit upper limits on human-robot contact forces: in the case of unexpected contact, the allowable pressure on different parts of the human body ranges from about 150N for the hand to about 35N for the abdomen. Traditional industrial robotic hands, due to the rigidity caused by high gear ratios, can hardly meet these requirements and must rely on physical isolation with safety fences. Robotic hands with structural compliance, on the other hand, are aligned with safety standards from their design principles, which will greatly simplify their certification processes in markets such as the EU and the US—at a time when the regulatory pathway for humanoid robots to enter homes remains unclear, meeting the strictest safety standards in advance is a strategic first move to secure early market approval.

In terms of durability, the complete finger assembly has been validated through millions of cycle tests, with a lifespan sufficient to support long-term commercial deployment.

Deep Vertical Integration and Mass-Production Capability

Perhaps what best embodies 1X's ambition is its layout at the manufacturing level. The company has achieved deep vertical integration: the motors, custom electronic components, and tendon systems are all self-developed.

The Business Logic of the Vertical Integration Strategy: Vertical integration refers to a company bringing multiple links of the supply chain under its own control, with Apple designing its own chips and Tesla building its own battery factories being typical examples. In the robotics hardware field, this strategy is especially critical. Before the 2020s, robotics companies generally relied on Japanese and German suppliers for key components: Harmonic Drive's precision harmonic reducers, Bosch Rexroth's servo drives, and Heidenhain's high-precision encoders. This outsourcing model was viable in the small-batch research prototype stage, but when the target is annual production of tens of thousands or even hundreds of thousands of units, single-point supply-chain dependencies and high component costs become insurmountable obstacles—if the cost of just a humanoid robot's motors and reducers reaches tens of thousands of dollars, the whole-unit price can never enter the mass market. The Tesla Optimus project explicitly lists in-house development and production of motors and actuators as a core strategy, and companies like Figure and Agility are also progressively internalizing key components. 1X's self-developed motors, custom electronics, and tendon systems—while simultaneously opening up the potential option of supplying other humanoid robot manufacturers—is a typical competitive strategy of "manufacturing as a moat."

The Exponential Challenge of Manufacturing Yield: Going from hundreds to ten thousand units is not merely a linear increase in quantity but an exponential leap in manufacturing system complexity. The yield rate of every component in a dexterous hand must be extremely high to guarantee whole-unit yield—if a hand has 25 joints, each involving 5 key parts, then even with a per-part yield as high as 99%, the whole-hand yield would only be about (0.99)^125 ≈ 29%, meaning over 70% of products are scrapped. This "yield multiplication effect" is a core challenge in precision mechanical manufacturing, and is why the yield management systems of the aerospace and semiconductor industries are so complex and expensive. If 1X can truly achieve its mass-production goal of ten thousand units per year, it means it has established a systematic quality assurance system in process control, incoming material inspection, and assembly standardization—rather than relying on manual fine-tuning. Building this manufacturing capability is often harder, more time-consuming, and more difficult for competitors to quickly replicate than technical R&D.

Hundreds Already Built, with an Annual Capacity Target of Ten Thousand

According to reports, 1X has already manufactured hundreds of robotic hands and has the capacity to produce 10,000 this year. At a time when humanoid robots generally remain at the prototype stage, this mass-production capability stands out. It means 1X leads not only technologically but has also completed its preparations in supply chain and scaling—which is often the key factor determining who can truly bring products to commercial reality.

The Arms Race for Dexterous Hands Has Begun

Across five dimensions—degree-of-freedom design, transmission architecture, sensing capability, safety mechanisms, and manufacturing scale—1X's robotic hand offers quite convincing answers. It reflects a clear trend in humanoid robot competition: whoever can build a hand that is both dexterous and safe, and can also be mass-produced, holds the key to real-world application scenarios.

The Gap Between Technical Demonstration and Actual Deployment: The robotics industry has a well-known pattern that researchers call the "Lab-to-Real Valley of Death": systems that perform excellently in controlled lab environments often see performance plummet in unstructured real-world environments. The core factors driving this phenomenon include: out-of-distribution perception—the lighting, cluttered backgrounds, and unseen object types of real environments cause vision systems to fail; contact uncertainty—the surface texture, mass distribution, and friction coefficients of real objects are far more complex than lab props; and long-tail failure modes—occasional edge cases during long-term operation continuously accumulate. For 1X, between the spec announcement and "stably completing tasks in cluttered real-world environments" still lies the need for extensive real-scenario data accumulation and targeted reinforcement learning training. This is also why mass-production capability is so critical—only large-scale deployment can generate sufficiently diverse real interaction data to drive model iteration into a positive cycle.

Of course, there is still a gap between specs and demonstrations, and the real test lies in whether these hands can stably complete tasks in cluttered real-world environments. But judging from this release, 1X has pushed the "hand"—the hardest component of humanoid robots to master—to the critical threshold of engineering and commercialization.