TL;DR
A data-driven framework that optimizes robot hand morphology using large-scale human manipulation motion data.
Problem
Jointly optimizing robot body design and control is a huge combinatorial challenge. Existing work focuses on learning control, but automatic generation of hand structures from data is underexplored.
Approach
Collect over 4 million frames of human fingertip motion from everyday tasks.
Use the same simple inverse-kinematics control policy for both design evaluation and post-fabrication control.
Optimize tree-structured robot hands to reproduce target motions.
Train an RL actor to propose good designs and joint angles, reducing search time from hours to minutes.
Fabricate designs as one-piece articulated structures with print-in-place joints.
Results & Contribution
The 6-DoF hand achieves better teleoperated fingertip tracking than commercial robot hands.
Specialized 3-DoF hands reproduce structured trajectories with reduced mechanical complexity.
Demonstrates that large-scale human motion data can optimize not only controllers but also physical robot embodiments.