A Two-Stage Optimization Framework for Designing Direct-Drive Dexterous Robotic Hands

Abstract

The emergence of micro-motors has made it possible to design direct-drive dexterous hands with flexible in-hand manipulation capabilities and without complex mechanical transmission errors. However, existing design methods in this context still rely on heuristic trial-and-error, lacking systematic methodologies. We propose a two-stage optimization framework that formalizes the design of direct-drive dexterous hands as an automated, gradient-free optimization problem. Using the Covariance Matrix Adaptation Evolution Strategy (CMA-ES), the framework autonomously resolves highly-coupled design parameters under the physical constraints of selected micromotors. The first stage identifies an optimal hand configuration by maximizing the opposability envelope volume. After that, the second stage refines phalanx proportions by reproducing natural human hand motions within a physics simulator. By optimizing a composite loss function that strictly penalizes physical unfeasibility, the optimized design reduces self-collision depth by 50.49% and joint limit violations by 32.01% compared with the empirical baseline. This methodology connects theoretical kinematics with physical hardware constraints, rapidly generating optimal dexterous hand design from micro-motors.

Zhenyuan Zhang

Dexterous hand design

Mingqi Chen

Postdoctoral Researcher

Qiang Li

Professor and header of AG

My research interests include underwater robot, collaborative robots, humanoid robots.