An Actuator-aware Genetic Algorithm for Soft Robot Design
Abstract
Soft robot design involves the collaborative optimization of morphology and control. Particularly, genetic algorithm is popular for morphology optimization and reinforcement learning is commonly adopted for control. However, many existing genetic algorithms adopt mutation only without crossover, ignoring the recombination advantage of high-quality solutions. Traditional crossover operators may cause performance degeneration since they ignore the functional relationships between adjacent voxels in soft robots. To address this issue, this paper proposes a new morphology-similarity-based parent selection strategy and a new actuator-aware crossover operator to generate high-quality offspring. Particularly, two similar parents are selected from the population based on their Hamming distance in the decision space to reduce the structural damage in the recombination process. On the two selected parents, the local structure of actuators with adjacent voxels are taken as an unit and exchanged to generate new candidate solutions. Integrating the parent selection and crossover into the action inheritance-based evolutionary algorithm, a new algorithm named actuator-aware genetic algorithm (AGA) is put forward. Fourteen tasks on the EvoGym platform are taken to verify the performance of the proposed algorithm. The experimental results show that AGA achieves better performance in most tasks, especially in complex tasks such as Hurdler-v0 and GapJumper-v0.