Xiaomi unveils 21-DOF biomimetic dexterous hand after testing 1,827 factory SOPs
Xiaomi disclosed a 21-DOF biomimetic dexterous hand designed around flexible manufacturing at its car factory, after analyzing 1,827 SOPs; it says 17-21 DOF outperformed 6-15 DOF by 15%-38% in stable grasp success.
Simulation ablation analysis based on real operation data showed that in the 17-21 degree-of-freedom (DOF) range, the average stable grasp success rate of the biomimetic hand was 15%-38% higher than that of hands with 6-15 DOF, Xiaomi said. The 21-DOF design delivered the largest improvement in factory operation coverage and grasp success rate. In this scenario, raising the DOF to 23 or above produced a clear slowdown in grasp improvement, reducing the marginal benefit of adding DOF, according to the company.
Based on those findings, Xiaomi designed a new 21-DOF biomimetic dexterous hand. It uses a hybrid actuation and tendon tension control scheme to achieve precise grasp force control and improve reliability and maintainability, while the company also systematically optimized the cooling system and full-palm tactile sensing.
Chen Duan, director of the Digital Economy Integration and Innovation Development Center at Central University of Finance and Economics and deputy director of the Intelligent Products and Industry Working Committee of the Chinese Association for Artificial Intelligence, said Xiaomi's approach of taking real factory operations as the design starting point and letting scenarios define the product is a more durable moat than hardware parameters. Tactile, force-position and assembly-timing data collected on the shop floor are scarce scenario-based assets, he said; they can be used to train the hand's operation policies and continuously feed Xiaomi's AI system, improving cross-scenario transfer capability.
Xiaomi presented the findings as technical exploration results, with the hand designed around factory scenarios. The account focused on the relationship between DOF and factory grasp performance, the hand's actuation and sensing design, and the use of factory data in robot training.