Learn and evolve to optimize robot morphologies

It is challenging for robots to complete tasks in dangerous environments. Traditional approaches require redesigning and rebuilding specialized robots for each specific task, leading to inefficiency. Inspired by the diverse animal morphologies shaped by evolution, the study introduces a novel two-layer evolutionary method for generating robot morphologies. In the external evolution stage, morphological structural changes are carried out by selecting the robots with high fitness in the population. In the internal evolution phase, each robot adjusts its morphological attributes in each episode before interacting with the environment. Our approach ensures that robots can quickly adapt to environmental changes through dynamic adjustment of morphological properties while maintaining morphological diversity through changes in morphological structure. The central architectural element of this strategy is Graph Neural Networks (GNNs) and the parameters of the strategy are updated by Proximal Policy Optimization (PPO). Experimental results across four different environments demonstrate the effectiveness of the proposed method. Robots evolved through this method exhibit higher performance in terms of morphological design compared to baseline methods.