Design and Control Co-Optimization for Automated Design Iteration of Dexterous Anthropomorphic Soft Robotic Hands

TL;DR

The paper tackles automated design iteration for dexterous anthropomorphic soft Robotic hands by co-optimizing hand design and control policies in simulation using genetic algorithms and policy transfer. It fabricates top designs via 3D printing and validates them through teleoperation on a standardized six-object task suite, showing that simulation-driven rankings align with real-world performance and that optimized designs surpass prior hands. Across 18 simulated object instances and six real-world objects, two designs (v6 and v7) emerge as top performers, closely tracking real-world outcomes despite sim-to-real gaps. The work demonstrates the practical viability of simulation-guided design for soft hands and highlights future directions for broader design spaces and direct policy deployment.

Abstract

We automate soft robotic hand design iteration by co-optimizing design and control policy for dexterous manipulation skills in simulation. Our design iteration pipeline combines genetic algorithms and policy transfer to learn control policies for nearly 400 hand designs, testing grasp quality under external force disturbances. We validate the optimized designs in the real world through teleoperation of pickup and reorient manipulation tasks. Our real world evaluation, from over 900 teleoperated tasks, shows that the trend in design performance in simulation resembles that of the real world. Furthermore, we show that optimized hand designs from our approach outperform existing soft robot hands from prior work in the real world. The results highlight the usefulness of simulation in guiding parameter choices for anthropomorphic soft robotic hand systems, and the effectiveness of our automated design iteration approach, despite the sim-to-real gap.

Figures (5)

• Figure 1: Our approach includes: a) hand design optimization in simulation using genetic algorithms and policy transfer; b) CAD for 3D-printing optimized hand that outperforms other hand designs in simulation; and c) real-world design evaluation of optimized hand using teleoperation on the same set of manipulation tasks.
• Figure 2: a) CAD of top-performing optimized hand v7 with DIP, PIP, and MCP joints labelled, as well as tendon placements along the finger shown in blue. The distal, middle, and proximal phalanges are also labelled in green to show part of the hand design parameters. b) Visualization of simulated top-performing hand design v7.
• Figure 3: Top three optimized hands. From left to right, the success rate AUC (%) of the hands averaged over all 18 object instances in simulation experiments are 53.32, 53.07 and 52.47 respectively.
• Figure 4: The 6 goal poses (shown for pen object) used for real world teleoperated manipulation tasks.
• Figure 5: (Top) simulation evaluation results averaged for pick up and reorient tasks for each of the six objects at scale $1\times$ on randomized goal poses with AUC success rate given by Equation \ref{['eq:evaluation']}; and (Bottom) Real-world teleoperation evaluation averaged on same objects for six goal poses using grasp quality metric explained in Section \ref{['sec:teleop_setup']}.