Experts Talk Robotics
Just one week after releasing the bipedal humanoid robot G1+, Unitree Robotics has launched its next-generation dexterous hand, the DEX5-S. Such a release cadence measured in "weeks" is hard to ignore.
01. DEX5-S: Two Configurations for the 22-DOF Hand
The core parameters of the DEX5-S are clear: 22 degrees of freedom (DOF) per hand, designed in a 1:1 scale with a real human hand (102×187×26mm), weighing 620 grams in total, featuring an aluminum alloy shell, and starting at CNY 39,900. The DOF distribution is 5 for the thumb, 4 each for the index/middle/ring fingers, and 5 for the pinky. Compared to the previous generation, it adds a roll DOF to the pinky, significantly improving the overall grasping dexterity.
It is available in two models, the Standard Edition and the Pro Edition, with basically identical hardware bodies: all 22 motors support direct drive, backdrivability, and dual-encoder control. The maximum payload per hand is 2kg, the continuous operating payload is 1kg, it supports 15V–65V wide-voltage power supply, and the peak power is 200W.
Source: Unitree Robotics
The differences between the two versions focus on the perception level—the Standard Edition only features basic feedback for position, torque, and IMU, while tactile perception is exclusive to the Pro Edition, targeting advanced physical AI research and supporting secondary development of tactile grasping algorithms.
There are three noteworthy details regarding the technical breakthroughs of this new product.
First, each joint features a built-in extreme impact torque protection structure to protect the gears from damage under external impacts, while ensuring motion accuracy through precision micro-reducers. Second, the product has undergone bending life tests, with the official emphasizing "ultra-long bending life," directly addressing the durability pain points of dexterous hands in high-frequency training scenarios. Third, under a Gigabit Ethernet configuration, the control frequency reaches up to 1000Hz, and the palm camera can be directly connected to the dexterous hand without external wiring, a design that significantly lowers the threshold for whole-machine integration and deployment.
02. Where Does 22 DOF Stand?
According to statistics from the "2026 China Dexterous Hand Industry Report" released by Experts Talk Robotics, over the past year, dexterous hands with 20+ active DOFs have emerged in batches. With the addition of the DEX5-S released by Unitree this time, a group of high-DOF dexterous hands is presented more intuitively.
—This list continues to grow.
The position of the DEX5-S in this lineup can be evaluated across three dimensions.
In terms of DOF, it stands in the mid-to-upper tier of the first echelon. The 22 DOFs are on par with the main hand DOFs of the OmniHand 3 Ultra-T, higher than the 21 of the XHAND 1 PRO, and have a slight advantage over the 20 of the O30 and Wuji Hand 2. However, the gap is already very small, so small that it is hard to constitute a substantial functional barrier.
In terms of pricing, the DEX5-S exhibits a clear "entry-level pricing" characteristic. The starting price of CNY 39,900 is at the affordable end among 20+ DOF products. For comparison, a single Shadow Dexterous Hand costs about USD 110,000, the JD.com price of the Revo 3 is close to CNY 200,000, and the procurement unit price of the XHAND 1 PRO is about CNY 25,500, though there are information discrepancies in the labeling of active DOFs. The DEX5-S further pushes down the price anchor for 20+ DOF dexterous hands. Coupled with Unitree's own H1/G1 main body ecosystem, it is actually reducing the comprehensive cost for developers to access high-DOF dexterous manipulation capabilities.
In terms of perception, the DEX5-S adopts a "version-split" strategy, which is precisely the real focus of current industry competition. The Standard Edition strips away tactile perception, which is only equipped in the Pro Edition—this means Unitree positions tactile perception as an "advanced capability" rather than a "basic configuration." However, the industry trend is exactly the opposite: tactile perception is shifting from an optional configuration to a standard capability. The Revo 3 comes standard with a full-palm tactile array and fingertip visuo-tactile fusion sensors; the XHAND 1 PRO is equipped with 18 distributed tactile sensors covering the fingertips, finger pads, and palm; and the OmniHand 3 Ultra-T features a visuo-tactile fusion perception system and an in-palm camera. The tactile solution of the DEX5-S Pro is not outstanding in terms of parameters. This could be an intentional trade-off in Unitree's current product cadence, or it could be reserved space for subsequent iterations.
03. From "Scarce Capability" to "Entry Ticket"
Before 2025, 20+ DOF was still an exclusive label for a few products. The Shadow Dexterous Hand, with its 20 active DOFs, occupied a benchmark position in the scientific research market for over fifteen years. By 2026, however, there are more than ten named 20+ DOF dexterous hands. The highest DOF in the industry has reached 42, surpassing the 27 DOFs of a human hand. As the marginal returns of "stacking DOFs" diminish, the focus of competition will inevitably shift.
The direction of this shift is clear: from a hardware parameter race to a triangular competition involving control algorithms, tactile perception, and data closed loops.
In terms of control algorithms, the concentrated emergence of fully direct-drive solutions is the result of algorithms driving hardware. The motion model of direct drive is more deterministic and predictable, possessing a natural advantage in Sim2Real transfer training—policies trained in simulation environments can be deployed more smoothly onto real robots, significantly reducing adaptation costs. This means that the "algorithm friendliness" of hardware design is becoming a new competitive edge.
In terms of tactile perception, the industry is tackling the challenge of "grasping firmly and touching accurately." Xense Robotics' VTLA embodied tactile model integrates vision, language, touch, action, and proprioceptive states into a single system, backed by a dataset containing 1,000 hours of real physical interaction records. Tashan Technology, on the other hand, approaches this from the chip level, launching a dynamic tactile perception chip that directly converts tactile signals into pulse signals. Tactile perception is no longer just a matter of "how many sensors are installed," but rather "how tactile signals enter the model's decision-making chain."
In terms of the data closed loop, this is the most decisive link. The industry currently does not lack "hands," nor does it lack prototypes; what it lacks are products that can land and work in real scenarios, and whoever can first run through the positive cycle of "technology-scenario-data." The tactile data, force feedback data, and motion trajectory data generated by dexterous hands in real operations are the core fuel for training the next generation of manipulation policies. Whoever masters the channel for high-quality data collection takes the initiative in iteration speed.
04. What Has Unitree Done Across These Three Dimensions?
Between control algorithms and the data closed loop, Unitree has chosen a more ambitious entry point: open-sourcing an embodied foundation model, making the dexterous hand an "execution terminal" for the model rather than isolated hardware.
On September 10, Unitree fully open-sourced the UnifoLM-WLA-1.0 general humanoid foundation model. With 6B parameters, it was trained on approximately 2,500 hours of high-quality real-robot data. A single model coordinates 64 tasks (54 desktop manipulation tasks + 10 whole-body manipulation tasks), covering scenarios such as taking out the trash, making the bed, tidying the kitchen and bathroom, folding clothes, and storing plates. It is not bound to a single end-effector—it can drive both parallel grippers and two types of dexterous hands.
This means that from its inception, the DEX5-S is embedded in a model ecosystem that has already run through 64 tasks. Developers who get the DEX5-S can perform scenario adaptation and fine-tuning based on UnifoLM-WLA-1.0. The significance of the open-source strategy lies in transforming Unitree from "selling hands" to "selling a platform for hands"—open-sourcing the model lowers the entry threshold, which brings in more developers and more scenario data, and the scenario data, in turn, feeds back into the model's generalization capabilities. This is a positive cycle that is now being initiated.
In addition, the trade-offs in the integrated design of the DEX5-S also serve the data collection scenario. The design where the palm camera is directly connected to the dexterous hand under a Gigabit Ethernet configuration without external wiring reduces the complexity of whole-machine wiring. For data collection tasks that require long-term, high-frequency operation, the improvement in reliability itself is a guarantee of data quality.
The strategy of exclusive tactile perception for the Pro Edition can be understood as a pragmatic data strategy: first, use the Standard Edition to rapidly expand the deployment base at a lower price, accumulating manipulation data at the position and torque levels; meanwhile, use the Pro Edition to target advanced researchers willing to pay for tactile data, maintaining a presence in the track of tactile perception which has not yet converged. Running two versions in parallel ensures shipment volume without betting too early on a tactile technology route that is not yet clear.
Final Words
The DEX5-S itself is a balanced product: 22 DOFs, real-hand size, fully direct-drive and backdrivable, starting at CNY 39,900—placed in the 2026 market, it is not breathtaking, but solid enough.
High DOF is no longer the moat for dexterous hands. The competition for dexterous hands is shifting towards the generalization capability of control algorithms, the data quality of tactile perception, and the engineering closed-loop capability to continuously extract training signals from real scenarios.
This article is compiled based on publicly available online information and is for reference only, not constituting investment advice.