Gasgoo Munich-For the past two years, the race in dexterous hands has been a numbers game: 6, 12, 22 degrees of freedom, with research versions pushing as high as 42. The implicit assumption seems to be that the higher the number, the closer the machine gets to a human hand.
Yet there's an awkward disconnect: robots capable of dancing or making heart shapes on stage often struggle back in the real world. They might fail to twist open a soda can, let alone gently hold a raw egg.
Having more joints doesn't mean knowing how to work.
That's why Xiaomi is taking the opposite approach this time.
On October 27, Xiaomi unveiled the core specs of its new bionic hand: 21 degrees of freedom, a hybrid drive system with 26 motors, and high-density tactile sensing across 1,500 points on the palm.
Yet the company didn't lead with the number "21." Instead, it started with a story: those 21 degrees of freedom were "calculated" from 1,827 standard operating procedures (SOPs) inside its own car factory.
Image source: Xiaomi Technology
21 Degrees of Freedom, Derived from 1,827 Factory SOPs
To understand Xiaomi's logic, you first have to see where the industry is stuck.
Degrees of freedom—essentially the number of directions a finger can move independently—is the most intuitive spec for a dexterous hand, and the easiest to weaponize for marketing. But in reality, labeling isn't standardized: some cite "total degrees of freedom," others "active degrees of freedom," while passive or underactuated linked joints often get lumped into the count.
Between 6 and 42, the numbers might look impressive, but the actual capabilities can vary wildly.
More importantly, degrees of freedom don't come for free. Every extra active joint requires stuffing another set of drives, transmissions, and sensors into the palm. That adds weight, heat, control complexity, and potential points of failure.
And this hand happens to be one of the most expensive components in the entire machine. Morgan Stanley estimates that the dexterous hand accounts for roughly 17.3% of the Tesla Optimus's bill of materials.
So the question becomes: how many joints does a hand actually need for factory work? Too few, and it can't grip or pinch accurately; too many, and you're paying for flexibility you don't use while sacrificing reliability.
You can't answer this by gut feeling. Whether you aim high or low, you need evidence.
Xiaomi's method flips the script: instead of designing the hand first, it started by watching how human workers use their hands in the factory.
Specifically, Xiaomi's technical team selected 16 typical work sections and 2,579 standard operating procedures at the Xiaomi EV manufacturing pilot center. They performed multi-angle video and motion capture on 1,827 of those procedures, covering over 70.8% of manual tasks. The captured data was then classified, segmented, cleaned, and redirected for multi-dimensional analysis and simulation based on human hand movements.
Put simply, it's like running "time-and-motion studies" on the hands of frontline workers to see exactly how much each joint is actually being used.
The findings came in two layers, and they're quite revealing.
Image source: Xiaomi Technology
The first layer: human hand joints don't really "slack off."
Xiaomi measured utilization by dividing the range of motion during actual tasks by the average full stroke of that degree of freedom. It found that across the covered tasks, the utilization rate of nearly every tested degree of freedom exceeded 50%.
This suggests these joints aren't sitting idle during work, but whether they "can be merged" ultimately depends on coupling and simulation ablation studies—utilization rates alone aren't enough to decide.
The second layer: human hand joints have natural coupling; they don't all work independently.
For instance, on the same finger, the flexion angle of the distal interphalangeal joint (DIP) is typically only one-third that of the proximal interphalangeal joint (PIP), and the two are highly correlated during movement.
In other words, two independent active degrees of freedom on a human hand actually move in tandem, meaning a single drive source could replace them. This is a classic example of underactuated design.
Following this logic, Xiaomi's team calculated the correlation between degrees of freedom across joints using factory data. This established a biomimetic basis for underactuation: if the correlation between two joints exceeds a certain threshold (such as 0.8), they are considered candidates for underactuated coupling design.
Utilization and coupling provide a general direction for the feasible range, but pinpointing the specific boundaries depends on simulation design and engineering constraints. Xiaomi hasn't disclosed whether these boundaries were set a priori or derived from data fitting.
What ultimately locked in this range was simulation ablation testing.
Xiaomi's team simulated hands with varying degrees of freedom repeatedly grasping objects. The results showed that the 17–21 degree-of-freedom range achieved an average stable grasp success rate 15% to 38% higher than the 6–15 range. Notably, 21 degrees of freedom offered the greatest improvement in both factory task coverage and grasp success rates.
However, increasing to 23 or more degrees of freedom resulted in significantly slower gains in grasping ability, marking a decline in the marginal returns of stacking degrees of freedom.
That is how Xiaomi arrived at 21 degrees of freedom for its new dexterous hand. It's not about showing off; it's a balance point calculated from its own factory scenarios and simulation tests.
This approach actually aligns with a growing consensus in the industry.
Zhou Chen, co-founder and CEO of Dexterous Intelligence, once noted that embodied intelligence operates differently from the "hardware-first, software-later" logic of cars and industrial robots—it follows a "model-defined hardware" approach.
Xiaomi, however, pushes the starting point of that "definition" one step further: using real-world motion data to determine hardware specifications first.
But data can only answer "how many degrees of freedom are needed." It can't answer the other question: how exactly do you cram those degrees of freedom into a hand the size of a palm?
Setting the Degrees of Freedom Doesn't Guarantee the Hand Can Be Built
Integrating 21 degrees of freedom into a palm-sized space creates immediate engineering conflicts.
First, there's weight and layout. The most direct approach for high degrees of freedom is to stuff a motor into every joint, but that turns the palm into a heavy, hot block of iron.
Xiaomi's new bionic hand primarily uses tendon drive, supplemented by 5 direct-drive degrees of freedom placed on joints with lower load requirements. This strikes a balance between degrees of freedom, load capacity, weight, and range of motion.
The cable-driven motors are relocated to the forearm, bringing the palm weight down to 324 grams. A lighter tip means lower rotational inertia, allowing the hand to react faster.
It's worth noting here: the palm of a cable-driven hand is light because the motors have been "shifted" to the forearm, not because weight vanished. When evaluating the whole system, the motors and weight in the forearm must still be counted.
This is also why the tendon route, despite being considered more biomimetic and suitable for high degrees of freedom, has never dominated the industry.
Moreover, tendons have inherent weaknesses: ropes creep and loosen over time, and once tension changes, control accuracy suffers.
Zhu Linghua, president and CTO of Dexterous Intelligence, put it bluntly: cable drive technology isn't new—elevators and cranes use it—but "the engineering implementation is incredibly difficult." It involves not just friction loss, preload stability, and system integration complexity, but also challenges in design, mass production processes, and reliability.
To compensate for the nonlinear control issues caused by tendon creep, Xiaomi developed a proprietary tendon tension sensor. It compensates for this nonlinearity in real time, maintaining long-term consistency and reliability in the cable drive system.
However, there is no data yet on how much this targeted solution can extend the system's lifespan.
Image source: Xiaomi Technology
Then there is the issue of heat, which cannot be ignored.
Just how hot can a hand get? Xiaomi cites figures showing that under harsh conditions, power consumption for a single hand can exceed 100 W. That's because gripping often involves holding a position for long, non-cyclical periods—essentially "holding your breath" to maintain an action—which results in sustained high current.
To put 100 W in perspective: that's higher than the peak power of many fast-charging smartphones.
To address this, Xiaomi developed a thermal prediction model. It claims to forecast temperature distribution and trends for the next 5 to 10 minutes, estimate the remaining time before overheating, and proactively implement thermal management measures like task scheduling, reducing joint torque, and adjusting fan speeds.
Xiaomi's team simultaneously explored two technical routes for this model: physical mechanism modeling and machine learning with online correction. Currently, the prediction accuracy is within 3°C.
Given that Xiaomi's robot features a full-body fabric cover—which adds to aesthetics and friendliness but also complicates cooling—the company is researching new heat-dissipating fabrics. It is exploring high-performance solutions through high thermal conductivity fillers, surface radiation modification, and natural moisture wicking and evaporation.
This might sound like a "soft" area, but for a hand intended to work on production lines for long periods, it's a critical "hard" metric for continuous operation.
Finally, there is tactile sensing.
Xiaomi's new bionic hand inherits the design philosophy of its previous tactile glove. It features 1,500 sensing points across the entire palm, with a tactile spatial resolution of 1 mm at the fingertips, finger pads, and finger bases, and 2 mm at the palm. The sensors are also replaceable and easy to maintain.
The function of these tactile points works on two levels. In the short term, they allow the hand to perceive contact position and force distribution during gripping, knowing exactly how much force to apply and whether an object is slipping.
In the long run, this contact data—structurally identical to human hand contact—serves as core corpus for training models.
The Hype Around Dexterous Hands Continues, But It's Not Time to "Hand in the Exam" Yet
Looking beyond the technology, it's worth weighing the significance of this development.
The dexterous hand sector is certainly not lacking in hype right now.
According to data from IT Juzi, financing in this sector exceeded 25 billion yuan in just the first half of 2026. Among the players, Zhidian Robotics, spun off from Agibot, has surpassed a post-money valuation of $1 billion, while Lingxin Qiaoshou, a specialist in robotic hands, is targeting a valuation of around $6 billion.
Yet beneath the flow of capital, the industry hasn't even settled on a unified technical path: tendon drives are biomimetic but difficult; linkages are stable but clumsy; and direct drives are precise but weak on torque.
Although Tesla's third-generation Optimus also relocated motors to the arm and uses a hybrid lead screw and tendon drive, this is more a choice of a direction with a "higher ceiling." Tesla itself uses a hybrid solution, which doesn't mean it's the final answer.
In this "out-of-focus" phase, the significance of Xiaomi's hand may not lie in proving which route is right, but in offering a methodology that isn't led by spec sheets: first clarify exactly what the hand is doing in the target scenario, then decide how many degrees of freedom to give it, what drive to use, and how to handle heat and sensing.
Of course, this hand is still far from "final exam day," and three boundaries need to be clarified.
First, the number 21 was calculated specifically for scenarios like Xiaomi's own car factory. If you switch to folding clothes at home or plugging cables on a consumer electronics assembly line, the distribution of actions changes, and the optimal balance point might no longer be 21.
Moreover, there is disagreement within the industry about whether a high-degree-of-freedom five-finger dexterous hand is even worth it in industrial settings. A Goldman Sachs survey in May showed that most manufacturers currently prefer a wheeled chassis with a 2-3 finger gripper, believing this setup can cover 70% to 90% of industrial applications.
In other words, the 70.8% coverage applies only to Xiaomi's own work sections, not the entire manufacturing sector.
Second, Xiaomi has not released data on the lifespan and cost—the two aspects most criticized for tendon drives. Even with wear-resistant materials like composite tungsten wires, costs in the industry remain in the tens of thousands of yuan range. Whether mass production and maintenance can hold up is a longer-term test than grasp success rates.
However, there are differing voices on whether "lifespan should be the priority right now." Du Dalong, founder and CEO of Octopus Power, believes that at this stage, there's no need to obsess over lifespan first. Instead, the focus should be on pushing the upper limits of capability. As long as the "brain" can drive the hand to perform valuable tasks, lifespan issues will be resolved in subsequent iterations as the supply chain matures.
Third, this is a technical exploration. There is no timeline for mass production, cost figures, or external supply availability. The hand is still in its "internship" phase and shouldn't be read as a market-ready product.
But the direction is clear. The competition in dexterous hands is shifting from "who has the most joints" to "who understands best how the hand actually moves while working."
42 degrees of freedom isn't necessarily the answer, and 21 isn't necessarily the finish line.
What's truly worth remembering is that this time, someone finally stopped rushing to update the numbers and instead went to the factory to count, one by one, more than a thousand real-world actions.
