Xiaomi Unveils New-Generation Bionic Hand: 21 Degrees of Freedom Derived from 1,827 Factory SOPs

In the past two years, the degrees of freedom of dexterous hands have been piled higher and higher: 6, 12, 22, and the highest research version has already reached 42. It is as if the higher the number, the closer it gets to a human hand. But the awkward thing is that robots that can dance and make finger hearts on stage may, back in the real world, be unable to twist open an unopened can of soda, and pinching a raw egg is even harder than climbing to the sky. Having many movable joints does not mean being able to do work. So Xiaomi's approach this time is the reverse. On October 27, Xiaomi announced the core information of its new-generation bionic hand: 21 degrees of freedom, hybrid drive with 26 motors, and 1,500-point high-density full-palm touch sensing. But Xiaomi did not throw out the number "21" first. Instead, it first explained one thing: these 21 degrees of freedom were "calculated" from 1,827 standard operating procedures (SOPs) in its own car factory. 21 degrees of freedom, derived from 1,827 factory SOPs To understand why Xiaomi did this, one must first look at where this industry is stuck. Degrees of freedom, simply put, are the number of directions in which the fingers can move independently. They are the most intuitive parameter of a dexterous hand and are also the easiest to use as a promotional number. But in reality, different companies label degrees of freedom inconsistently: some mark "total degrees of freedom," some mark "active degrees of freedom," and passive and underactuated linked joints are also counted in. Between 6 and 42, the numbers look lively, but the actions they can actually perform differ greatly. More importantly, degrees of freedom do not come for free. For every additional joint that can actively move, another set of drive, transmission, and sensing must be crammed into the palm, and weight, heat, control complexity, and failure points all rise accordingly. And this hand is also one of the most expensive components in the entire machine. Morgan Stanley has estimated that the dexterous hand accounts for about 17.3% of the material cost of Tesla's Optimus. So the question becomes: how many joints does a hand actually need for work in a factory? Make too few, and it cannot grasp or pinch accurately; make too many, and you are paying for flexibility you do not use, while also sacrificing reliability. This question cannot be answered by guesswork. Whether doing more or less, there must be evidence. Xiaomi's method is to reverse the order: do not design the hand first; first go see how people use their hands in the factory. Specifically, Xiaomi's technical team selected 16 typical work sections and 2,579 standard operating procedures (SOPs) at the pilot test center of Xiaomi Auto's intelligent manufacturing department, and collected multi-angle video and motion capture data for 1,827 of them, covering more than 70.8% of manual processes. The collected motion capture data were classified, segmented, cleaned, and retargeted, and then multidimensional data analysis and simulation based on the human hand were carried out. Put plainly, this is doing a "time study" for the hands of frontline workers, to see how much each joint is actually used when doing work. The conclusion has two layers, and they are quite interesting. The first layer is: human hand joints actually do not slack off much. Xiaomi used "range of motion in actual work ÷ the average full range of that degree of freedom in the human hand" to measure the utilization rate of each degree of freedom, and found that in the work covered by the data, the utilization rate of almost all measured degrees of freedom exceeded 50%. This shows that these joints are not idle in actual work, but whether they "can be merged" ultimately depends on coupling and simulation ablation; utilization rate alone is not enough to settle it. The second layer is: human hand joints have natural coupling; not every one works independently. For example, on the same finger, the bending angle of the distal phalanx (DIP) is usually only one-third that of the middle phalanx (PIP), and the two are highly correlated in movement. In other words, two independent active degrees of freedom are actually linked in the human hand and can be replaced by one drive source. This is a typical underactuated design. Following this line of thought, the Xiaomi team further calculated the correlation of degrees of freedom among joints based on the data collected in the factory, and derived a bionic basis for underactuation: if the correlation between two joints exceeds a certain threshold (such as 0.8), they can be considered to have the potential for underactuated coupling design. Utilization rate and coupling give the general direction of the feasible range, but the specific truncation of the interval boundaries still depends on the design of simulation experiments and engineering constraints. Xiaomi has not disclosed whether these boundaries were set a priori or are the result of data fitting. What really settled this interval was simulation ablation experiments. The Xiaomi team put hands with different degrees of freedom into a simulation environment to repeatedly grasp, and found that in the 17–21 degrees-of-freedom tier, the average stable grasp success rate was 15%–38% higher than in the 6–15 tier, and among them 21 degrees of freedom showed the greatest improvement in factory operation coverage and grasp success rate. But when increased further to 23 or more, the improvement in grasping ability clearly slowed, and the marginal benefit of piling on degrees of freedom declined. Xiaomi's new-generation dexterous hand's 21 degrees of freedom came from this: not showing off, but a balance point calculated in Xiaomi's own factory scenarios and simulation tests. This approach actually lands on an industry consensus. Zhou Chen, co-founder and CEO of Lingqiao Intelligence, said something: embodied intelligence is different from the logic of automobiles and industrial robots, which is "hardware first, software later"; it is "models define hardware." However, Xiaomi pushed the starting point of "definition" one step further: first use real motion data to set the hardware specifications. But data can only answer "how many degrees of freedom should be made"; it cannot answer another question: how exactly should these degrees of freedom be crammed into a palm-sized hand. Degrees of freedom are set, but that does not mean the hand can be made Once the 21 degrees of freedom are set, integrating them into a palm-sized space immediately brings engineering contradictions. First is weight and layout. The most direct way to achieve high degrees of freedom is to stuff one motor into each joint, but then the palm would become a heavy, hot lump of iron. Xiaomi's new-generation bionic hand primarily adopts tendon-driven actuation, combined with 5 direct-drive degrees of freedom placed on joints with low load requirements, striking a balance among degrees of freedom, load capacity, weight and range of motion. The tendon-drive motors are moved to the forearm, bringing the palm weight down to 324 grams. With a lighter end effector, the moment of inertia is smaller, so the hand responds faster. One more point here: the tendon-driven hand's palm is light because the motors have been "transferred" to the forearm, not because weight was reduced out of thin air. When evaluating the whole machine, the motors and weight in the forearm still have to be counted. This is also why the tendon route, although considered more bionic and better suited to high degrees of freedom, has never managed to unify the field. Moreover, tendons themselves have weaknesses: ropes creep and slacken over time, and once tension changes, control becomes inaccurate. Lingqiao Intelligence President and CTO Zhu Haohua once put it bluntly: tendon-drive technology is nothing new—elevators and cranes use it—but "engineering implementation is extremely difficult," involving not only friction losses, preload stability and system integration complexity, but also multiple challenges such as design difficulty, mass-production processes and reliability. To compensate for the nonlinear control problems caused by tendon creep, Xiaomi developed its own tendon tension sensor, which compensates for this nonlinearity in real time and maintains the long-term consistency and reliability of tendon-driven actuation. But how long a service life this targeted solution can achieve—there is currently no data. There is also the heat issue, which cannot be ignored. How hot can one hand get? The figure Xiaomi gives is that under harsh operating conditions, power consumption for a single hand can exceed 100 W—because grasping often involves holding a posture for long periods, non-periodically, "bracing with sustained effort," with current remaining high. What does 100W mean? It is higher than the peak power of many fast-charging phones. To address this, Xiaomi built a thermal prediction model, claiming it can predict the temperature distribution and trends for the next 5–10 minutes, estimate the remaining time before overheating, and carry out thermal management measures in advance such as task scheduling, reducing joint torque and adjusting fan speed. The Xiaomi team explored two technical routes for the thermal prediction model in parallel: physical mechanism modeling and machine-learning online correction, with prediction accuracy currently better than 3°C. Considering that Xiaomi's robot uses a full-body fabric-covered design, which enhances aesthetics and approachability but also places higher demands on heat dissipation, Xiaomi is also researching new heat-dissipating fabrics, exploring high-performance heat-dissipating textiles from multiple directions including high-thermal-conductivity material filling, surface radiation modification and natural moisture absorption and evaporation. This part sounds "soft," but for a hand that has to work on a production line for long periods, it is precisely a hard metric of whether it can operate continuously. Finally, touch. Xiaomi's new-generation bionic hand inherits the design approach of the previous generation's tactile glove, with 1,500 points covering the entire palm, a tactile spatial resolution of 1 mm between the fingers, on the fingertips and at the base of the fingers, and 2 mm on the palm, and it is replaceable and easy to maintain. The role of these tactile points can be divided into two layers: in the near term, they let the hand perceive contact position and force distribution during grasping, knowing how much force to use and whether an object is slipping. In the longer term, these contact data, isomorphic to the human hand, are also core corpus for training models. Dexterous hands keep drawing heat, but it is not yet time to "hand in the exam paper" After looking at the technology, step back and look at the significance of this matter. The dexterous hand track is not lacking in heat right now. According to IT Juzi statistics, in the first half of 2026 alone, financing in this track exceeded 25 billion yuan, among which Lingji Dian, spun off from Zhiyuan, has a post-investment valuation exceeding $1 billion, and Lingxin Qiaoshou, which specializes in hands, has a target valuation of about $6 billion. But beneath the hot money, the industry has not even finished arguing over a unified technical route: tendon-driven is bionic but difficult, linkage-driven is stable but clumsy, direct-drive is precise but weak in torque. Although Tesla's third-generation Optimus also moves motors to the arm and adopts a screw + tendon hybrid transmission, this is more a choice of a direction with a "higher ceiling," and Tesla itself is also using a hybrid solution, which does not mean it is the endgame. At this stage of "loss of focus," the demonstrative significance of Xiaomi's hand may lie not in proving which route is right, but in offering an approach that is not led by the spec sheet: first figure out what the hand is actually doing in the target scenario, then decide how many degrees of freedom to give the hand, what actuation to use, and how to dissipate heat and sense. Of course, this hand is still some distance from "handing in the exam paper," and three boundaries need to be made clear. First, the number 21 was calculated for scenarios such as Xiaomi's own automotive factory. Switch to folding clothes at home or plugging ribbon cables on a 3C production line, and the motion distribution changes; the optimal balance point may no longer be 21. Moreover, whether a high-DOF five-fingered dexterous hand is really worth it in industrial scenarios is itself a matter of disagreement in the industry. A Goldman Sachs survey in May this year showed that most manufacturers currently prefer wheeled chassis plus 2–3-finger grippers, believing this solution can cover 70%–90% of industrial applications. In other words, the 70.8% coverage applies only to its own work sections, not to all of manufacturing. Second, Xiaomi has not disclosed the lifespan and cost that the tendon solution is most criticized for. In the industry, even wear-resistant materials such as composite tungsten-wire tendons still cost on the order of ten thousand yuan, and whether mass production and maintenance can hold up is a longer-term test than grasp success rate. However, there are also different voices in the industry on "whether lifespan should be the first thing to compete on now." Du Dalong, founder and CEO of Octopus Dynamics, believes that at this stage there is no need to compete on lifespan first; instead, raise the capability ceiling first. As long as the brain can drive the hand to complete valuable operations, the lifespan issue will be resolved in subsequent iterations and as the supply chain matures. Third, this is a technical exploration article, with no mass-production timeline, no cost, and no timetable for external supply. The palm is still in an "internship" stage and should not be read as a product launch. But the direction is clear. Competition in dexterous hands is shifting from "whose hand has more joints" to "who better understands how the hand actually moves when doing work." 42 degrees of freedom may not be the answer, and 21 may not be the endpoint either. What is truly worth remembering is that this time, finally, someone did not rush to refresh the numbers, but first went to the factory and counted more than a thousand real motions one by one.