Phased genetic algorithm for simultaneous morphology and controller optimization in soft robots

Le résumé fourni par la source

The tight coupling between morphology, material distribution, actuator layout, and control makes soft robot design a high-dimensional and expensive optimization problem when body and controller are co-designed. In this paper, “co‑design” refers specifically to morphology evolution with per‑candidate reinforcement learning of control, not to the joint genetic evolution of brain and body. We propose a phased genetic algorithm (GA) that evolves voxel‑based morphologies; each candidate is evaluated by training an independent PPO controller under a fixed budget. Three key mechanisms are introduced: 1) a task-aware prior initialization that injects weakly biased, physically meaningful morphology templates into the initial population; 2) a large-block crossover operator (exchanging 2 $$\times$$ 2 or 3 $$\times$$ 3 voxel blocks) that preserves functional morphology modules; and 3) a phased evolutionary schedule that balances exploration and stabilization. Compared with four baselines, the proposed method achieves competitive overall performance, with clear advantages on Climber-v0 and Thrower-v0 and faster convergence on Walker-v0. Not applicable.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.