Zhiyuan Robotics split core businesses for independent fundraising, focusing on dexterous hands and embodied data to crack humanoid robot tech bottlenecks.
The humanoid robotics industry is spending hundreds of millions of dollars in R&D trying to teach robots something extremely basic: fine manipulation with their hands.
Picking up a strawberry without crushing it and placing it in a basket is a skill that human infants take months to learn and industrial robotic arms will never master. The degrees of freedom, force feedback, and precision control involved form a hard technical barrier running through the entire supply chain.
Zhiyuan Robotics' recent moves look somewhat unconventional. Instead of following the old path of "bigger and stronger," it has spun out its core business units, allowing each to raise funds independently, hire its own talent, and develop on its own. Dexterous hands, data platforms, quadruped robots, and leasing — one company has split into a small ecosystem.
To understand this, you first need to understand where humanoid robotics is currently stuck.
Dexterous Hands: The Hard Engineering Wall of Robotic Fine Manipulation
The hand is the hardest segment of the humanoid robotics technology chain.
The human hand has 27 degrees of freedom. Fingertips can perceive texture changes at the 0.1-millimeter level, and grip strength can shift from gentle to firm in an extremely short time. Industrial grippers only solve grasping and sacrifice everything else. To build a robotic hand that truly works, you need to pack three systems — actuation, sensing, and control — into a space of a few dozen millimeters, and all three must communicate and respond in real time.
The number of joints is the first hurdle.
Common two-finger grippers on industrial robots provide one or two degrees of motion and can handle standardized pick-and-place and assembly on a production line. But real-world tasks are far more complex: tightening a screw requires precise torque from the fingers, folding clothes requires multi-finger coordination, and handing medicine to an elderly person requires continuously sensing the pill's weight to adjust grip. To increase the number of joints, each additional joint means one more actuator, one more sensing and control loop, and the volume, weight, and power consumption of the entire hand grow accordingly — yet the available space for a dexterous hand is extremely limited.
Currently, there are three mainstream technical approaches, each with trade-offs.
The first is tendon-driven actuation. The principle is to use thin cables or steel wires to pull finger joints, allowing motors to be placed closer to the torso along the arm, which keeps the hand lighter and thinner. Tesla's Optimus robot used this approach in its early versions.
But tendon-driven systems face two unavoidable problems: first, the cables themselves fatigue and stretch under repeated tension, causing drift in control precision; second, multiple tendons passing through the same joint interfere with each other, making the overall control model extremely complex.
The second approach is direct drive. Each joint houses a small motor that directly drives the joint's rotation, offering fast response, strong torque controllability, and easily guaranteed position accuracy. The challenge is that fitting enough motors into the palm space significantly increases size and weight.
The third approach is pneumatic or hydraulic actuation. These use air or fluid pressure to drive artificial muscles or soft structures, providing natural grasping and gentle contact forces. But air pumps and hydraulic pumps are bulky and noisy, making integration into a mobile robot platform an engineering challenge in itself.
None of the three approaches is a perfect answer. Tendon-driven is lightweight but hard to tune; direct drive is precise but hard to slim down; pneumatic is soft but hard to integrate.
Force feedback is another hurdle.
A robot finger must be able to sense in real time how much force it is applying. The human hand's tactile feedback loop has a latency of about 20 milliseconds — an order of magnitude shorter than the time it takes to blink. Currently, the latency of this loop in most robotic dexterous hands is far higher.
When latency is too high, grasping fragile objects becomes unstable.
The sensor solutions themselves are also a constraint. Strain gauges offer high precision but limited sampling rates; piezoelectric films respond quickly but have noisy signals; fiber-optic sensing is difficult to package. With limited fingertip surface area, no single solution currently dominates for laying down a sufficiently dense sensing array.
Precision control is the third hurdle.
When all five fingers move simultaneously, each finger affects the force distribution of the others. Engineering approaches go in two directions: one is building accurate physical models for feedforward compensation, but friction, wear, and temperature changes cause the model to drift continuously; the other relies on learning, but the data requirements are extremely high.
Zhiyuan spun out its dexterous hand business into Lingjiedian (Critical Point), which became independent in January 2026 and completed three rounds of funding within a month, led by Tencent. The customer base for dexterous hands extends beyond humanoid robots: industrial exoskeletons, medical rehabilitation, and collaborative robots all have demand, and the market boundary is far larger than that of complete humanoid units.
Embodied Data: A Bottleneck Harder Than Algorithms
Training large language models requires decades of accumulated text data from the internet. Training models that teach robots to "do things with their hands in the real world" requires action data of robots operating in the physical world — and that kind of data barely exists on the internet.
This is the hardest bottleneck for the entire industry to bypass.
One usable training data point requires a robot to complete a specific task while recording multimodal information such as vision, touch, and joint angles, followed by manual annotation of action intent. Research from Stanford University shows that collecting 50 high-quality demonstration data points of dual-arm manipulation takes roughly a full day of manual operation.
When collecting data, how to record the human operator's action intent is a question without a standard answer. For the same task of "moving a cup from table A to table B," different operators may show vastly different hand trajectories, grasping styles, and force application rhythms. Some habitually slide the cup to the table edge before lifting; others lift vertically first and then move it horizontally.
To humans, these differences are all "correct answers," but when mixed together and fed into a model, the trained policy develops mode confusion.
Annotation itself is another challenge. Text data annotation is relatively objective, while physical operation data annotation involves a great deal of implicit judgment: Is this motion part of the "exploration phase" or "execution phase"? Is this frame's force change intentional or an operational error?
Annotators need a background in robotics, and such talent is scarce.
Diversity coverage of data is the third hurdle. To cover scenario variations, the same task must be collected repeatedly across different backgrounds, lighting conditions, and operator hand characteristics. And as deployment environments expand, new scenario data must be continuously added.
Zhiyuan spun off its data business into Mifei Technology (FindBee), with Sequoia leading the investment in February 2026 (source: STAR Market Daily). If Mifei opens up the data collection, annotation, and management infrastructure from within Zhiyuan to the entire industry, it can serve competitors like Unitree, Fourier, and CloudMinds, and its business boundaries won't be limited by Zhiyuan's own complete unit shipments.
Simulation Platform: The Technical Trump Card Kept at the Parent Company
Since real data collection is so expensive, replacing it with simulation data is a natural idea.
The advantages of simulation environments are obvious. In a virtual world, a robot can complete in seconds a training task that would take minutes in reality. Dozens of virtual environments can run in parallel, scenes can be randomly generated with no upper limit, and the risk of breaking hardware is zero.
Genie Sim 3.0 is Zhiyuan's simulation platform, which has been open-sourced. Notably, unlike Lingjiedian and Mifei, this line was not spun out; it stays with the parent company. The simulation platform is the technical foundation, deeply tied to the main model and data training on the robot itself, and spinning it out would create fragmentation. Zhiyuan has spun out businesses that can stand alone and raise their own funds, while keeping the technologies that must interlock with the robot body in-house. The boundary is drawn quite clearly.
But simulation data has an old problem known in the industry as the simulation-to-real gap. When a robot learns something in the virtual world, it often doesn't work well when transferred to the real world. Friction coefficients, object weights, and material elasticity in the simulation environment are all approximations. Each individual error may seem small, but in the real world these factors compound simultaneously, and the errors accumulate layer by layer.
The industry currently has three main approaches to reducing this gap.
The first is domain randomization training. During training, the various physical parameters in the simulation environment are deliberately perturbed at random, so the policy is not an optimal solution for one fixed parameter set but has some adaptability to a range of variations. This method works well for tasks requiring coarse-grained robustness, but if the perturbation range is too small, it doesn't cover real-world variation, and if it's too large, the training signal becomes chaotic.
The second approach is hybrid training. Large amounts of simulation data build the foundational capability, and then small amounts of real data fine-tune it. This approach has proven successful in the language model domain.
The challenge for embodied tasks is that the distributional bias of simulation data may be more systematic than in the text domain. If the amount of real data is insufficient during fine-tuning, the bias introduced by simulation cannot be fully corrected.
The third approach is making the simulator itself more realistic. One direction is using rendering-level techniques like neural radiance fields to scan real scenes and generate high-fidelity virtual environments. Another is introducing more accurate contact mechanics models into the physics engine.
The latter is more difficult because there is a fundamental contradiction between the computational cost of accurate physical simulation and the speed required for real-time training.
Open-sourcing Genie Sim is a two-way move. Outwardly, it attracts developers to contribute scenes and task definitions; inwardly, it builds a user base. But the long-term value of a simulation platform ultimately depends on how well the policies perform when deployed on real robots.
60 Robots Enter the Expo Hall — the Site Is the Exam
During WAIC, Zhiyuan's subsidiary Qingtianzu deployed 60 full-size humanoid robots across three venues and four halls, performing public-area navigation, retail, interaction, and entertainment tasks with bilingual Chinese-English support. According to The Paper's report on July 17, no one in the industry had previously done this at a major expo.
Having 60 robots online simultaneously is an engineering problem of an entirely different magnitude from single-unit lab testing.
Dispatch is the first hard threshold. Sixty robots online at once requires a real-time task allocation system. Visitor flow in the halls changes dynamically — during peak hours, allocation density must increase automatically; during off-peak hours, robots must be dispatched to charging and standby.
Communications is the second threshold. Each robot continuously uploads sensor data and downloads control commands. The expo halls are also full of other wireless devices competing for bandwidth.
Once the network jitters, the robots' real-time obstacle avoidance and task execution are affected.
Multi-robot collaborative obstacle avoidance is the third threshold. Sixty robots moving simultaneously in a limited space each have limited field of view, and they need to share location information to avoid blocking each other.
The hardest thing to predict in advance is visitor behavior. Expo visitors do unpredictable things to robots: suddenly crouching down to block their path, asking questions in mixed Mandarin and English, or reaching into a robot's operation space. These edge cases are nearly impossible to enumerate in the lab.
Qingtianzu's business logic therefore carries a hidden layer of value: commercial deployment itself is a data collection channel. Every day a robot operates in a real-world scenario, it generates real interaction data that cannot be mass-produced in a lab.
Final Thoughts
Among the four hard technical walls — dexterous hands, data, simulation, and scenario deployment — Zhiyuan has spun out the three that can stand alone as businesses, keeping simulation in the parent company as the technical foundation. Each corresponds to a confirmed hard problem in the industry.
According to an Omdia report, Zhiyuan's shipments are expected to exceed 5,000 units in 2025, compared to roughly 4,700 for Unitree in the same period. Shipment volume is the story of today; dexterous hand precision, data supply, simulation transferability, and scenario reliability are the variables that will determine who is still at the table two years from now.
After the split, each subsidiary must find its own way to survive and cannot rely solely on the parent company's capital. Lingjiedian must sell dexterous hands externally; Mifei must sell data services externally; Qingtianzu must recover costs through commercial deployment; and the quadruped line, Zhiyuan Kute (Kutuo), must carve out a position in Unitree's territory. The parent company holds two technical trump cards: Genie Sim and the main model. The pace and pressure across the four lines are completely different from when they were all under one roof.
For humanoid robots to reach mass production and deployment in real factories or homes, there are still several hard segments left to crack in the technology chain. Zhiyuan has chosen to split itself apart to tackle them separately — that's one path, not the only answer.