Introduction
China's dexterous-hand demonstrations are becoming harder to compare from videos alone. At the September 2026 Global Digital Trade Expo, Xynova showed the hybrid tendon/direct-drive Flex 2 and direct-drive Prima 1. INSPIRE showed a 24-DOF hybrid hand at the World Robot Conference and later used its hands in the World Humanoid Robot Games opening ceremony.
The practical problem is what happens after first contact. Fingers can cover the camera's view, an object may begin to slip and a controller can crush a delicate item while trying to stop the drop. Useful evidence has to include contact sensing, force behavior, cycle time and repeatable task results.
Source review date recorded for this article November 11, 2026.
Key findings
- Gasgoo reports Flex 2 as a 23-DOF, 400 g hybrid-drive hand with a 12 kg peak grasp load and a 4 kg continuous operating load; the manufacturer separately describes 19 active and four passive degrees of freedom. These are reported product figures, not a common independent benchmark.
- Gasgoo reports Prima 1 with 22 DOF, a claimed 0.08-second opening/closing cycle and 0.5 N force-control precision. The official Xynova website gives other metrics such as a 0.2 N minimum backdrive force and 0.5 mm repeatability, which measure different properties.
- An IEEE Robotics and Automation Society report describes a tactile-control comparison across five contact-rich tasks: a studied system averaged 62.8% success versus 28.2% without tactile input. That research result cannot be assigned to the Xynova or INSPIRE products.
How to compare dexterous hands without treating DOF as a success score
| Measure | Test setup | What to report |
|---|---|---|
| Actuation | Count active vs passive joints and specify drives | Controlled axes and usable grasp postures |
| Contact feedback | Same object material and occluded grip | Contact timing, force error and slip recovery |
| Holding force | Peak test and sustained test at stated wrist pose | Duration, load and temperature |
| Dexterous task | Cable insertion or object reorientation from changed poses | Completed tasks, damage and human retries |
| Durability | Repeated cycles and recalibration intervals | Sensor drift, wear and maintenance |
| Policy transfer | Train on one sensor/hand; test on another | Success change and extra tuning examples |
Two Xynova hand designs solve different actuation problems
Xynova's Flex 2 combines tendons and direct drive. At Hangzhou's September trade expo, Gasgoo listed 23 mechanical degrees of freedom, 400 g mass, a 12 kg peak grasp load and a 4 kg continuous load. The 2026 World Robot Conference exhibitor description breaks the 23 DOF into 19 active and four passive degrees. Passive motion can help a hand conform to an object, but it is not another independently commanded motor.
Prima 1 is presented as a direct-drive research hand with 22 DOF. Gasgoo reports 0.5 N force-control precision and a 0.08-second opening/closing time. Xynova's own product page separately lists 0.2 N minimum backdrive force, repeatability within 0.5 mm and a 1 kHz response bandwidth. These describe force interaction, return-to-position behavior and control response respectively; they are not interchangeable measures of grasp success.
The comparison needs a specified finger configuration, object mass and test duration. A peak-load figure may describe a brief event while the continuous load concerns longer use. A supplier quote should state how long the grip lasted, what posture the hand used and whether the load was supported by friction or mechanical enclosure.
Problem 1: vision loses the contact point when fingers close
A camera can locate a cable before grasping, but fingertips often occlude the plug once the grasp is formed. Inserting a connector then requires feedback about first contact, lateral force and the moment the plug starts to jam. A small deformation or slip may not be visible in time.
Tactile cells or fingertip sensors can detect contact onset and force distribution. A controller can respond by slowing the wrist, reducing commanded force or reorienting the object instead of blindly replaying the original trajectory. For testing, record contact force, insertion depth, jamming and whether the connector is damaged over multiple starting offsets.
A sensor's mere presence is not enough: the tactile stream must be time-aligned with wrist motion and read quickly enough for the contact dynamics. Otherwise, the correction can arrive after the fingers have already slid off the object.
Problem 2: a hand that closes strongly may still crush or drop objects
Grip strength and careful handling pull in different directions. A flexible bag can slip under a weak grasp; a thin plastic package can crease or tear under excessive force. The useful control variable is the margin between the minimum force needed to prevent slip and the force that damages the item.
A practical test uses different objects including a smooth bottle, a compliant pouch and a small rigid connector. Record dropped or damaged objects and the number of corrective grasps. Repeat after the fingertip surface has accumulated wear and after calibration, because sensor drift and finger friction alter contact estimates.
For a fair comparison keep the same object poses, wrist path and assistance rules across hands. Give separate results for initial pick, stable hold, in-hand repositioning and final placement instead of reporting only a successful video take.
Problem 3: touch data collected with one hand may not transfer to another
Different hands sense touch through resistive pads, optical gels, force sensors or other hardware. Outputs vary in physical units, coordinate frames, spatial resolution and sampling rate. A model trained only on one sensor can misinterpret another sensor's values unless they are calibrated and mapped to a common representation.
The September 2026 IEEE Robotics and Automation Society review describes research combining more than 3,000 hours of tactile records across 21 sensor types. It also discusses an experiment in which emphasizing unexpected tactile signals improved average success across five contact-rich tasks to 62.8%, versus 28.2% for a comparison model without touch. Those numbers belong to their respective studies and do not rank commercial hands.
A transferable dataset should record sensor type, calibration, fingertip contact geometry, time alignment, object material and failure labels. When evaluating a new sensor, exclude it from training and test whether the policy still detects slip and regulates forces; report the robot and object setups used.
What INSPIRE's public demos reveal and what a buyer should measure
INSPIRE introduced its RH524J1 24-DOF tendon/direct-drive hybrid hand at the 2026 World Robot Conference. Its World Humanoid Robot Games opening-ceremony report describes lifting and placing an energy-transfer token. The sequence shows that a coordinated event maneuver was performed, but it does not supply a distribution of grip failures under changed object weights or positions.
For industrial purchasing, ask for end-to-end trials with cycle time, contact-force logs, fingertip replacement intervals and the fraction completed without human intervention. In a cable insertion task, compare successful fully seated connectors, bent pins and recovery attempts. For repeated bag handling, measure object damage and lost grips after a long run.