Voxel-Based Soft Robot Co-Design

Motivation

Designing soft robots requires simultaneously optimizing both their physical structure (morphology) and control policy — a computationally expensive process. Traditional CPU-based approaches cannot scale to the massive evaluations needed for evolutionary optimization.

Approach

Built two GPU-accelerated frameworks for voxel-based soft robot co-design:

• Diff-CoDesign — Built on Taichi Lang, leveraging differentiable physics for gradient-based optimization
• VoxelCoDesign — Built on NVIDIA Warp, using evolutionary strategies with mass-spring physics simulation

Key technical contributions:

• Achieved 10.8x topology build speedup through GPU parallelization
• Batch-evaluated 200 robots per generation across 35 independent runs of 1000 generations each
• Completed 7 million total evaluations for comprehensive design space exploration
• Implemented reverse-mode automatic differentiation for physics-based gradient computation

Results

The frameworks demonstrate that GPU-accelerated co-design can make previously intractable optimization problems feasible, enabling thorough exploration of the joint morphology-control design space for soft robots.

Significance

This work bridges the gap between evolutionary robotics and modern GPU computing, showing that massively parallel simulation can unlock new possibilities in automated robot design.