Gait Generation for Quadruped Robots Using Deep Reinforcement Learning with Spike Representation

Abstract

Quadruped robots are promising mobile robots for practical applications due to their higher locomotor ability in adverse conditions such as construction and disaster sites. In recent years, previous studies have proposed many methods for controlling legged robots using deep reinforcement learning. Additionally, spiking neural networks (SNNs), which use spike representations of a neuron model, have gained attention due to their robustness against noise. In this study, we apply SNNs to deep reinforcement learning to generate the gait of a quadruped robot in a simulation. Moreover, we verified the robustness of the learned gait patterns against sensor noise.