Neuroevolutive Algorithms for Learning Gaits in Legged Robots

Gait generation for legged robots is a challenging task typically requiring either a hand-tuning design or a kinematic model of the robot morphology to compute the movements, generating a high computational and time efforts. Neuroevolution algorithms with the ability to learn network topologies, such as Neuroevolution of Augmenting Topologies (NEAT), Hypercube-based NEAT (HyperNEAT), and $\tau $ -HyperNEAT, have been used in the computational community to learn gaits in legged robots. An extended version of HyperNEAT called ES-HyperNEAT, where the substrate hosting the nodes of the neural network evolves in shape, seems to be a promising algorithm to evaluate gait learning tasks. Using two four-legged robot platforms with different degrees of freedom (Quadratot and ARGOv2), we compared the performance of a variety of neuroevolution algorithms based on HyperNEAT for learning gaits. The comparative analysis of the results reveals that the three evaluated algorithms, HyperNEAT, $\tau $ -HyperNEAT and ES-HyperNEAT, successfully generate gaits given a fitness function. In particular, for the Quadratot platform, ES-HyperNEAT learns faster and better than the other two methods, a result that is not observed in the ArgoV2. Additionally, ES-HyperNEAT produces more phase changes between joints movement, resulting in more natural robot movements. Finally, ES-HyperNEAT produces complex substrates from simple Compositional Pattern Producing Networks, CPPN-NEAT, allowing the simplification of the underlying neural network.