Xynova Flex 2
Last reviewed
Sources
14 citations
Review status
Source-backed
Revision
v2 · 2,694 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
14 citations
Review status
Source-backed
Revision
v2 · 2,694 words
Add missing citations, update stale details, or suggest a clearer explanation.
| Xynova Flex 2 | |
|---|---|
| Product information | |
| Manufacturer | Xynova (Hangzhou, China) |
| Public launch | May 13, 2026 |
| Generation | Second |
| Type | Hybrid-drive dexterous hand |
| Tagline | "Bionic Versatility, Intelligent Stability" |
| Total degrees of freedom | 23 (19 active, 4 passive) |
| Palm weight | 400 g |
| Repeatability | ±0.1 mm |
| Force control accuracy | 0.05 N |
| Single-hand grasp load | 12 kg |
| Rated continuous grasp load | 4 kg |
| Service life | About 1 million open-close cycles |
| Open and close speed | About 2 fist extensions per second |
| Drive architecture | Hybrid (cable-driven plus direct-drive joints) |
| Sensing | Multimodal (tactile, force, proximity, vision via wrist camera) |
| Control layer | "Cerebellum"-style adaptive grasping and slip detection |
| First public demo | ICRA 2026, Vienna, Austria, June 2, 2026 |
| Predecessor | Xynova Flex 1 |
The Xynova Flex 2 is the second-generation dexterous hand developed by Hangzhou-based robotics startup Xynova, publicly launched on May 13, 2026 as what the company calls the world's first bionic dexterous hand to use a hybrid "tendon rope plus direct motor drive" architecture. The hand offers 23 total degrees of freedom (19 active and 4 passive), a 400 g palm, ±0.1 mm positional repeatability, 0.05 N force control accuracy, a 12 kg single-hand grasp load, a 4 kg rated continuous operation load, and an approximately 1 million cycle service life, with multimodal sensing for adaptive grasping and slip detection. [1][2][3][12]
The Flex 2 succeeds the pure tendon-driven Flex 1 of August 2025 and is the first Xynova hand to combine cable-driven and direct-drive joints in a single end effector. Industry coverage framed the Flex 2 as a deliberate pivot from the Flex 1's raw load capacity toward sensing, precision, and control, and treated it as a serious candidate to become the hand of choice for outside humanoid integrators that prefer not to build their own. Xynova first demonstrated the physical hardware in public at the IEEE International Conference on Robotics and Automation (ICRA) 2026 in Vienna, Austria, on June 2, 2026. [3][4][5][12]
The launch came amid a rapid escalation of Xynova's funding. By the time of the product announcement, Xiaomi Strategic Investment had joined both the December 2025 angel round and the March 2026 Pre-A round, alongside CATL Capital, China Electronics Technology Group (CETC) affiliated funds, and others. In the weeks after the Flex 2 launch the company closed a Series A worth hundreds of millions of yuan, bringing cumulative funding to nearly 1 billion yuan (about US$148 million) raised in roughly its first 18 months, and was named one of the "New Eight Steeds of Hangzhou" emerging-technology companies on May 18, 2026. [3][12][13]
The Xynova Flex 2 is a humanoid robot end effector: a five-fingered robotic hand designed to replace a human hand on a humanoid robot or robotic arm and perform fine, force-sensitive manipulation. It is built by Xynova, a Hangzhou full-stack dexterous-manipulation company founded at the end of 2024 that also designs and manufactures the micro electric cylinders, hollow-cup motors, planetary roller screws, and joint modules inside its hands and sells those components to other robotics integrators. The Flex 2 is positioned for embodied AI systems that need a human-proportioned hand with closed-loop precision rather than maximum brute-force grip. [3][6][12]
The published specification sheet on Xynova's English product site as of May 2026 is summarised in the table below. Where a parameter has a Flex 1 counterpart, that figure is shown alongside for context. [1][2][6]
| Category | Parameter | Flex 2 value | Flex 1 value for comparison |
|---|---|---|---|
| Structure | Total degrees of freedom | 23 (19 active, 4 passive) | 25 (20 active + 5 passive) |
| Structure | Drive architecture | Hybrid: cable-driven plus direct-drive | Pure tendon-driven |
| Structure | Fingers | 5 (fully opposable thumb) | 5 (fully opposable thumb) |
| Structure | Kapandji thumb opposition score | 10 of 10 (perfect score) | Not separately published |
| Dimensions | Palm weight | 400 g | 380 g |
| Dimensions | Form factor | 1:1 adult human hand proportions | 1:1 adult human hand proportions |
| Load | Single-hand grasp load | 12 kg | More than 30 kg |
| Load | Rated continuous grasp load | 4 kg | Not separately published |
| Load | Fingertip force | About 20 N per finger | More than 20 N per finger |
| Durability | Service life | About 1 million open-close cycles | Not separately published |
| Precision | Repeatability | ±0.1 mm | Not separately published |
| Precision | Force control accuracy | 0.05 N | Not separately published |
| Dynamics | Open-close speed | About 2 fist extensions per second | Full palm cycle in about 0.6 s |
| Dynamics | Response time | Millisecond-scale dynamic response | Not separately published |
| Control | Back-drivable mechanism | Yes | Not separately published |
| Control | Control modes | Hybrid force and position control | Position control with tendon force feedback |
| Sensing | Tactile | Multi-region, palm and fingertip | Not separately published |
| Sensing | Force | Yes | Tendon-derived |
| Sensing | Proximity | Yes | Not separately published |
| Sensing | Vision | Wrist-mounted camera | Not separately published |
| Environment | Palm surface | Wear- and puncture-resistant | Not separately published |
| Environment | Ingress and drop rating | Dust- and drop-proof, advertised for extreme environments | Not separately published |
Several specification numbers benefit from unpacking. The ±0.1 mm repeatability and 0.05 N force control figures are the two metrics most often cited in third-party coverage, because they place the Flex 2 ahead of the published tendon-driven competition on precision while keeping the palm under half a kilogram. [1][2] The 23-DOF count, organised as 19 active and 4 passive degrees of freedom, is a reduction from the Flex 1's 25, but Xynova attributes this to consolidating passive degrees of freedom into the new hybrid drive geometry rather than reducing biomechanical coverage. [2][12] The reported perfect score on the Kapandji Thumb Opposition Test, a clinical assessment of thumb-to-fingertip and thumb-to-palm reach, indicates that the thumb workspace covers every point on the four other fingertips and into the palm. The roughly 1 million open-close cycle service life is the headline durability claim. [1][2]
The most significant change between the Flex 1 and the Flex 2 is the move from a pure tendon transmission to a hybrid architecture that mixes cable-driven joints with direct-drive joints in the same hand. The hybrid layout is intended to capture the compactness and biomimicry of tendons while removing the cable stretch, friction hysteresis, and backlash that limit the precision of pure-tendon designs. The result is a back-drivable mechanism that supports hybrid force and position control with ±0.1 mm positional repeatability and 0.05 N force control. [1][2][6]
The hybrid choice places the Flex 2 between the two main camps in the high-DOF dexterous hand market. Pure tendon-driven designs such as the Tesla Optimus Gen 3 hand, the Linkerbot LinkerHand L30, and the Shadow Robot Hand keep all actuators in the forearm and route cables through pulley networks to the finger joints. Pure direct-drive designs such as the Wuji Hand place all actuators inside the hand itself and dispense with tendons. The Flex 2's hybrid layout uses both, with the tendon-driven joints providing the biomechanical reach and the direct-drive joints providing the precision-critical motions. [1][2][6][7]
The drive system is built around Xynova's own micro electric cylinders, which deliver 100 to 300 N of thrust in a 10 to 12 mm package, and on its own hollow cup motors (8 mm diameter) and 7 mm planetary roller screws that convert rotary motion into the linear finger actuation. The actuator stack is vertically integrated, with motors, controllers, reducers, and screws all manufactured in-house. The same component portfolio is sold to other humanoid robot integrators as standalone parts. [3][6][14]
The Flex 2's other major change relative to the Flex 1 is a much richer sensing and control stack. Xynova describes the system as multimodal perception fusion plus a "cerebellum"-style control layer. [1][2][5]
The Flex 2 integrates multiple sensing modalities. Tactile sensors are distributed across the palm and fingertips for surface contact and slip detection. Force sensing is read from the actuators themselves through the back-drivable transmission. Proximity sensing supports pre-contact approach, and a wrist-mounted camera provides close-range visual perception of the manipulation workspace, positioned at the wrist joint rather than in the centre of the palm so that the camera is not occluded by the fingers during a grasp. [2][5][8][12]
The combination of force and tactile feedback is the primary input to the slip detection feature, which the company advertises as triggering compliant reflexes when an object begins to slip in the grasp. The wrist camera adds a visual channel for tasks where the object pose changes during the grasp, such as in-hand reorientation of a tool. [1][2]
Xynova uses the metaphor of a "cerebellum" to describe the dedicated control layer that sits between the perception stack and the joint controllers. The layer is responsible for adaptive grasping, slip detection, and compliant reflexes, and is presented as the component that converts the raw sensing data into low-level joint commands fast enough for millisecond-scale dynamic response. The exact algorithmic details, including whether the control layer uses learning-based policies, classical impedance control, or a mix, are not disclosed in the public materials as of May 2026. [1][5]
The Flex 2 is not a strict upgrade over the Flex 1. It is a repositioning of the product in the design space. The deltas can be summarised as follows. [2][5][6]
| Axis | Direction of change | Magnitude |
|---|---|---|
| Total DOF | Down | 25 to 23 (now 19 active + 4 passive) |
| Palm weight | Up | 380 g to 400 g (about 5 percent heavier) |
| Whole-hand grasp load | Down | More than 30 kg to 12 kg |
| Drive architecture | Changed | Pure tendon to hybrid (tendon plus direct drive) |
| Sensing | Added | Tactile, force, proximity, wrist camera |
| Precision | Added | ±0.1 mm repeatability, 0.05 N force control |
| Control modes | Added | Hybrid force and position control on back-drivable mechanism |
| Cerebellum-style control | Added | Adaptive grasping, slip detection, compliant reflexes |
| Environmental robustness | Added | Dust- and drop-proof, wear- and puncture-resistant palm |
In essence, the Flex 1 maximised mechanical envelope while the Flex 2 maximises closed-loop precision and intelligent grasping. The two products coexist in Xynova's catalogue rather than the Flex 2 replacing the Flex 1. [5][6]
The Flex 2 launch was first reported in English on May 13, 2026 by the RoboHub account on X, which framed the spec sheet as "sharp" and the Xiaomi cap-table presence as the headline strategic detail. The thread emphasised that Xynova is not just building a hand but the broader manipulation stack of micro linear actuators, joint modules, sensing, control, and developer-facing integration. [3]
The Aihola technical writeup characterised the Flex 2 as "a deliberate pivot toward sensing and control" and argued that the apparent regression in DOF count and grasp load was a chosen trade-off rather than a degradation. The piece also noted what is not yet disclosed publicly: pricing, customer identities, independent benchmarks, sensor counts and types, and failure data on the tendon ropes and the "anti-creep" mechanism. The Threads coverage from the AI Continuum highlighted the millisecond-scale response time and the 23-DOF biomimetic configuration. [5][8]
The CyberRobo X account, posting on May 12, 2026, was the first to describe the Flex 2 as a "hybrid-drive" hand that combines cable and direct drive, and noted the wrist-mounted camera position as a deliberate departure from palm-centre cameras. [9] Interesting Engineering reported that the Flex 2 "combines cable-driven tendons with direct-drive actuation, balancing flexibility, compactness, and precision," and that Xynova would benchmark the hand against global peers at ICRA 2026. [12]
The Flex 2 launch is part of a broader pattern in which the high-end of the Chinese dexterous hand market is professionalising in step with the humanoid robot market. By the first half of 2026, Inspire Robots, Linkerbot, Xynova, Wuji Tech, and DexRobot were typically named as the leading external suppliers, while Tesla, Xiaomi, Unitree, and AgiBot were continuing to develop captive hands inside their own humanoid programmes. [6][10]
Funding around the Flex 2 was unusually fast. Xynova closed an angel round of more than 100 million yuan in December 2025 led by CATL Capital, a Pre-A in March 2026, and a Series A worth hundreds of millions of yuan around the May 2026 launch, three rounds inside roughly half a year and a cumulative total approaching 1 billion yuan (about US$148 million). Investors across the rounds include CATL Capital, Xiaomi Strategic Investment, CETC-affiliated funds, and the strategic investment arm of electric-vehicle maker Li Auto. "Congratulations to Xynova on completing three rounds of financing within half a year," said Lou Yang, a partner at exclusive financial advisor Light Source Capital. On May 18, 2026 Xynova was named one of the "New Eight Steeds of Hangzhou" emerging-technology companies. [12][13][14]
The Xiaomi connection is the most-cited element of the strategic story. By the Flex 2 launch, Xiaomi Strategic Investment had been on the Xynova cap table for roughly five months through the angel round and had topped up in the March 2026 Pre-A. Xiaomi's own humanoid hand program for the CyberOne platform was advancing in parallel, with an April 2026 redesign that introduced full-palm tactile sensing and an evaporative "sweating" cooling channel. It is not publicly confirmed whether the Xiaomi humanoid hand uses Xynova technology, and Xynova has not been listed as a hand supplier to CyberOne in Xiaomi's own communications. The investment direction is therefore at the component layer rather than at the integrated hand. [3][11]