How Xiaomi’s CyberOne Bionic Hand Achieves Full‑Palm Tactile Sensing and 150k Grasp Cycles

Overview

In a prior test on an automotive self‑tapping‑nut station, the Xiaomi Robotics system operated continuously for 3 hours, achieving a dual‑side installation success rate of 90.2% and a line cadence of 76 seconds.

Full‑Palm Tactile

To raise the success rate toward 100%, the CyberOne bionic hand was redesigned: volume was reduced by ~60% (from 228 × 105 × 64 mm to 187 × 88 × 36 mm), degrees of freedom increased by 64%, and full‑palm tactile sensor area grew to 8200 mm². The hand now supports more than 150 k grasp‑cycle repetitions.

Two challenges were identified: (1) obtaining tactile perception when visual occlusion occurs, and (2) heavy reliance on inefficient tele‑operation for tactile data collection. The team introduced a tactile glove that captures fingertip, finger‑pad, and palm data directly from a human operator, enabling rapid construction of large‑scale operation datasets and reducing wear on the bionic hand.

Bionic Morphology

The bionic hand is engineered to match human hand proportions (1:1) and to emulate human kinematics, reachable space, drive capability, inertia distribution, and response, thereby minimizing simulation‑to‑real gaps.

Reliability

Initial trials showed component failures (tendon cords, springs, sleeves) before 10 k repetitions, even for simple operations. Through iterative design‑simulation‑testing cycles, each component’s durability was improved, ultimately achieving a stable 150 k‑cycle lifespan with consistent tactile data acquisition.

Cooling Exploration

Higher active freedom required more motors; a single motor tower can exceed 100 W, with over 30 W dissipated as heat at 70 % efficiency. To address overheating, a metal 3D‑printed liquid‑cooling channel circulates coolant via a micro‑pump to an evaporation zone. In tests, the system evaporates 0.5 mL of water per minute, providing roughly 10 W of active cooling.

Application Attempts

The goal is for the bionic hand to replicate human‑like grasp softness, coverage, reachable space, speed, and load capacity. Human grasp data were collected, integrated into simulation, and used with imitation and reinforcement‑learning strategies to train thousands of digital parts toward human‑like grasp poses. Laboratory experiments continue to explore fine‑grained tactile manipulation.

Open Source and Resources

Test videos, including the 61‑hour 150 k‑cycle grasp test, are available at https://cdn.cnbj1.fds.api.mi-img.com/robot-open-source/150000-cycles.mp4.

TacRefineNet: https://sites.google.com/view/tacrefinenet

Xiaomi‑Robotics‑0: https://github.com/XiaomiRobotics/Xiaomi-Robotics-0