Author | Ma Guangyu Editor | Li Yuchen
Xiaomi has made new progress in robotics.
In recent days, Xiaomi unveiled a bionic dexterous hand designed for industrial scenarios. This hand features 21 degrees of freedom (DoF), driven by a hybrid system of 26 motors, and its palm covers approximately 1,500 tactile sensing points.
According to reports, to design this hand, the Xiaomi Robotics team began working within automotive factories. They selected 1,827 standard operating procedures out of 2,579 from 16 typical work sections for data collection and analysis, ultimately determining the structure of the dexterous hand.
Notably, it has been only about seven months since Xiaomi announced its humanoid robot entering an automotive factory for an "internship" before revealing this dexterous hand.
In March of this year, Xiaomi robots began attempting tasks such as self-tapping nut installation in automotive factories. By July, Xiaomi disclosed that the success rate of bilateral operations at this station increased from 90.2% to 98%, while adding internship positions for center console side cover plates and bin folding.
In terms of product development, the efficiency of the Xiaomi Robotics team is top-tier. However, beyond the product itself, although robots can complete station tasks, there is still a long way to go before they truly take on production duties. After all, besides success rates, factors such as production takt time, long-term stability, maintenance costs, and whether robots can adapt to more stations are issues that need to be resolved.
Nevertheless, the dexterous hand remains one of the most critical components in the field of robotics.
In factories, for robots to perform these actions, they require more flexible structures, more accurate tactile and force feedback, and control capabilities that can handle on-site changes. From this perspective, the dexterous hand unveiled by Xiaomi this time is not just a technical advancement; it also reflects that its robot R&D is becoming closer to real-world working scenarios: letting robots enter factories and then redesigning robots based on tasks within those factories.
From launching the robot dog "CyberDog" in 2021, to introducing the humanoid robot "CyberOne" in 2022, and now entering automotive factories for internships, Xiaomi has spent five years traversing the path of robotics.
01CyberOne Internship
Xiaomi's entry into robotics began before this current wave of embodied intelligence.
In August 2021, Xiaomi CEO Lei Jun presented Xiaomi's first-generation bionic quadruped robot, CyberDog, nicknamed "Iron Egg," at a launch event. During the same event, Xiaomi announced the establishment of a robotics laboratory.
The first generation of Iron Egg released 1,000 engineering exploration units priced at CNY 9,999, primarily targeting developers and robotics enthusiasts. Its task was to help Xiaomi begin accumulating capabilities in robot bodies, joints, motion control, and perception.
A year later, Xiaomi launched its first-generation humanoid robot, CyberOne, also known as "Iron Big."
This robot stands approximately 1.77 meters tall, weighs 52 kilograms, and has 21 joint DoF across its body. For Xiaomi at the time, the difficulty was significant; simply keeping a humanoid robot balanced and enabling it to walk was already a demonstration of technical capability.
In 2023, Xiaomi introduced the second-generation robot dog, CyberDog 2, priced at CNY 12,999. Lei Jun candidly stated that this product was positioned as a development platform and was not suitable for ordinary users, implying it remained an experimental product.
From an external perspective, moving from quadrupeds to bipeds, Xiaomi continuously showcased new robot forms while gradually facing a more realistic question: Robots can be built, but what exactly can they do for users?
Xiang Diyun, General Manager of Xiaomi's Robot Business Unit, mentioned externally in mid-year that since the project was initiated in 2021, the internal team had been constantly considering the practical applications of robots. After 2023, Xiaomi rarely publicly disclosed humanoid robot progress, but R&D investment continued.
Over these years, Xiaomi continued to advance the mobility capabilities of its robots. Xiang Diyun noted that in 2024, a roughly 50-kilogram Xiaomi robot achieved running speeds of 12 to 13 kilometers per hour. However, regarding future development, Xiaomi gradually shifted its focus to how robots could complete specific tasks.
You Yangwei, Head of Bionic Control at Xiaomi's Robot Business Unit, once distinguished the development directions of robots in a conversation with Leifeng Network: one direction emphasizes high-dynamic movement capabilities such as running and jumping; the other requires entering factories to complete tasks like handling and assembly.
These two goals affect the overall machine design. Robots in factories must consider station height, arm reach, load capacity, and workspace, as well as energy consumption and cost. Robots with strong mobility can attempt factory tasks, but their hardware configuration may not be the most optimal choice.
In March of this year, Xiaomi's humanoid robot began its "internship" in an automotive factory.
According to disclosures at the time, the robot, based on the Vision-Language-Action model Xiaomi-Robotics-0 and combining multimodal perception with reinforcement learning, attempted tasks such as self-tapping nut installation and bin handling. At the March earnings call, Xiaomi Group President Lu Weibing introduced that the robot could operate autonomously for 3 hours, with an installation success rate exceeding 90%, meeting the fastest takt requirement of 76 seconds on the production line.
From this point, the pace of the Xiaomi Robotics team began to accelerate. In July, Xiaomi disclosed four months of internship progress: the bilateral operation success rate at the self-tapping nut installation station improved from 90.2% to 98%, while expanding to include center console side cover plates and bin folding scenarios.
At the World Robot Conference in August, the new generation of "Iron Big" made its public debut. This robot stands approximately 1.7 meters tall, weighs 66 kilograms, and has 66 DoF across its body.
Xiang Diyun explained that the new robot's height referenced the height of factory workers. Compared to the first-generation Iron Big, the new generation's dimensions were reduced to adapt to the actual working environment.
According to the Xiaomi team's introduction at the time, two robots were interning at different stations in the automotive factory: one was the previously developed high-dynamic robot, and the other was the new generation Iron Big. They undertook various tasks such as bin folding, part loading, and nut installation.
From the first-generation Iron Big to the new generation robot, Xiaomi's understanding of robot body size, hand capabilities, and task scenarios has evolved alongside the factory internship.
The robot's hand became a problem Xiaomi needed to solve during this process.
02Why Does Xiaomi Need a "Dexterous Hand"?
Xiaomi's automotive factory has a station for installing self-tapping nuts.
Workers need to remove nuts from feeding devices, adjust their orientation, place them into positioning fixtures, and subsequently have equipment assist in tightening them. During this process, the posture of each nut is not fixed, grasping must contend with magnetic interference, and the placement position and angle must be sufficiently accurate.
In March of this year, Xiaomi robots began interning at this station. After several months of training, their task success rate gradually improved, but compared to skilled workers, the gap was not only reflected in success rates.
In a conversation with Leifeng Network in July of this year, You Yangwei gave an example: At the nut installation station, a worker could grasp two parts with one hand and continuously complete placement; whereas the robot at that time mostly grasped one, placed one, then went to grasp the next, potentially requiring two robots to complete the same work.
Although this example referred to the situation at a specific station at that time and does not represent the configuration of all Xiaomi stations, it illustrates a key issue: A robot being able to complete a task and completing a task efficiently like a human are two different goals.
Building a pair of robot hands follows a completely different logic.
Picking up a part, unscrewing a bottle cap, or snapping a clip into a blind spot often requires no conscious thought about how each finger should move. But behind these actions lies continuous judgment of object shape, position, force, and contact state.
For robots to enter human working and living environments, they face similar problems. They must not only know where objects are but also judge from which position to grasp, with how much force, and whether the object moves after grasping.
Therefore, the dexterous hand is often regarded as the "crown jewel" of robot technology. It requires integrating structure, drive, tactile sensing, and control capabilities into a single hand, coordinating with the robot's vision, arms, and models to complete specific operations.
Xiaomi's starting point for designing this new generation dexterous hand was observing how workers use their hands. Among the 1,827 operating procedures collected in the automotive factory, the team observed the range of motion of hand joints and analyzed the kinematic relationships between different joints.
The human hand has many joints, but they do not always work independently. When performing an action, some joints need individual adjustment, while others follow the movement of other joints. For robots, if these correlations can be identified, it is possible to allow some joints to share drives, thereby reducing structural complexity.
More importantly, the Xiaomi Robotics team needed to know how many DoF a hand needs to cover sufficient operations.
Xiaomi conducted grasping simulations based on collected movements, tools, and parts in the factory. According to Xiaomi's disclosed results, under their test conditions, the average stable grasping success rate in the 17 to 21 DoF range improved by 15% to 38% compared to the 6 to 15 DoF range; further increasing to 23 or more DoF resulted in slowing improvements in grasping capability.
Balancing operational coverage and grasping performance, Xiaomi chose 21 DoF.
This hand adopts a hybrid solution primarily using tendon-driven mechanisms with some directly driven joints. Simply put, some motors are placed in the forearm, driving finger movement via tendons to reduce the burden on the palm. According to details disclosed by Xiaomi, the weight of the palm section is only 324 grams.
A lighter palm helps reduce the burden during end-effector movement, but the motors have not disappeared; the weight in the forearm still needs to be incorporated into the overall machine design. Tendons may experience tension changes after long-term use, affecting movement precision. Therefore, Xiaomi added tension sensors to the hand, hoping to improve control stability during grasping.
Regarding tactile sensing, Xiaomi arranged approximately 1,500 tactile sensing points on the fingertips, finger pads, finger roots, and palms of the dexterous hand. When the robot grasps an object, in addition to judging position through vision, it can utilize contact feedback to perceive the grasping state and further adjust the action.
However, from a physical structure perspective, the more complex the structure, the more engineering challenges arise. For example, multiple motors working together create heat dissipation pressure. Xiaomi disclosed that under extreme conditions, single-hand power consumption may exceed 100 watts. The team established a temperature prediction model, hoping to predict heating situations for the next few minutes in advance and adjust via task scheduling, torque, and fan speed.
But a more complex hand does not necessarily mean it is more valuable for all tasks.
For objects with regular shapes and fixed positions, grippers or dedicated equipment may suffice; some tasks requiring multi-finger operations may not necessarily need a high-DoF five-finger hand. Equipment procurement costs, maintenance difficulty, and long-term reliability all influence factory choices.
Previously, practitioners told Leifeng Network that dexterous hands with different DoF have applicable scenarios, and industrial clients ultimately look at task requirements, cost, and stability. The Xiaomi team has also made similar trade-offs.
They once attempted to have robots complete bin transport tasks from one location to another, choosing to temporarily skip this after verification. The reason was that such work might not require the hands and legs of a humanoid robot; other equipment could potentially complete it.
Xiaomi focuses more on tasks currently performed by workers, involving complex processes and requiring flexible operations, which traditional automation cannot easily cover in the short term.
This also means that Xiaomi's new hand must eventually answer a commercial question: Compared to simpler grippers, three-finger hands, or dedicated equipment, what additional work can 21 DoF accomplish, and can it offset the costs brought by a more complex structure?
Currently, Xiaomi has disclosed the design and testing process of the new hand but has not fully revealed when it will be installed on the full robot, when it will enter production lines, nor the long-term operational lifespan, maintenance costs, and mass production pricing.
These questions still need to be verified at specific stations.
03How Much Longer Is Xiaomi's Path to Embodied Intelligence?
After robots enter factories, the most direct criterion for validation is whether they can consistently complete work.
The 98% disclosed by Xiaomi is the bilateral operation success rate at a specific nut station. Public reports mention that the pass rate for skilled workers at the same station is approximately 99%. These two numbers appear close, but their specific statistical methods and comparison conditions have not been fully disclosed, so one cannot judge solely based on these two numbers that the robot is close to replacing workers.
Factories also need to consider production takt time, product pass rates, failure frequency, and manual intervention.
Different stations have varying tolerances for failure. Some line-side tasks can allow for retry time, while at stations following the assembly line flow, a single anomaly may affect subsequent takt times.
You Yangwei previously mentioned that for a robot to officially enter a station, it must simultaneously meet requirements for task success rate, product pass rate, and production takt time. Being able to perform the action is only the first step.
The on-site environment will not remain constant. Material racks may be moved, extra workers may appear nearby, parts may fall, and equipment status may change. Robots cannot only complete operations once under pre-set conditions; however, solving this problem relies on data.
Industrial data is not easy to obtain. Factories cannot stop production for extended periods to collect robot training data; building a separate complete production line for collection is too costly. The direction Xiaomi hopes to explore is allowing robots to acquire basic operational capabilities through large amounts of data first, then using relatively small amounts of on-site data to learn new station tasks.
In July of this year, Xiaomi released the embodied foundation model Xiaomi-Robotics-1. According to the paper, the model's pre-training used over 100,000 hours of real operation trajectories, with data collected via handheld grippers and cameras, covering home, commercial, industrial, and office environments.
In addition, Xiaomi released Xiaomi-Robotics-U0, exploring the generation of embodied scenarios and robot interaction data, hoping to reduce the cost of data expansion through generation and data augmentation. It differs from models responsible for generating robot actions, addressing the source of training data and how to increase scenario variations.
GPT-Image-2.0 and Xiaomi-Robotics-U0: Under the same instruction, GPT-Image-2.0 showed errors in robotic arm positioning.
This is a common challenge faced by the robotics industry and a hurdle Xiaomi must pass to take the next step.
From the robot dog in 2021, to humanoid robots entering factories, to the dexterous hand unveiled this time, Xiaomi has continuously invested in robotics for five years. During this period, Xiaomi's investment direction expanded from bodies, joints, and motion control to dexterous manipulation, data, and models, while the team continued to grow.
According to public information, Xiaomi plans to invest over CNY 200 billion in R&D within five years and has committed to investing over CNY 60 billion in the AI sector in the next three years. Large investments have led to team expansion; currently, the scale of Xiaomi's Robot Business Unit exceeds 200 people.
Robots require long-term investment in talent, hardware, data, computing power, and application scenarios, which may not bring direct returns in the short term. The core reason Xiaomi dares to invest is that it already possesses scenarios where robots can enter.
Automotive factories, smartphone manufacturing, and large appliance production all contain various types of operational tasks. Xiaomi can find stations within its own manufacturing system, allowing robots to enter real environments, observe problems, and repeatedly verify solutions.
Of course, proprietary factories also have production rhythms and cost requirements. Even when interning in its own factories, robots cannot indefinitely impact normal production. They still need to prove value through specific tasks.
Previously, at a communication meeting, the Xiaomi team stated that if only two or three fixed stations are solved, whether to use humanoid robots is not the most important factor. What the team hopes for is that after solving one station, robots can migrate existing capabilities to other stations more quickly. If every additional station requires redesigning hardware, recollecting large amounts of data, and retraining models, the application cost of robots will be difficult to lower.
If the same body, hand, and model can learn more tasks with appropriate adjustments, the general-purpose value of robots has the opportunity to gradually manifest. The 21-DoF dexterous hand unveiled this time is one of the new milestones.
Whether it can help robots complete work faster and more stably, and whether it can control costs and maintenance frequency during long-term operation, still needs answers from real stations. And these answers will determine how far Xiaomi's robots can go in the next step.