QuadNet: simultaneous online trajectory and footstep planner for quadrupedal robot using deep neural network

Abstract

We present a deep learning framework that generates the quadrupedal robot’s base body trajectory and footsteps simultaneously in a short time. Our model consists of three sub-networks: terrain encoder, state feature network, and trajectory generator. Our model receives the initial robot state, target goal base body pose, and terrain heightmap as input. The state feature network and terrain encoder extract features from these input data. These features guide the trajectory generator to produce precise motion for a given terrain. Experiments show that our motion planning approach is executed within 130ms, generating up to 7.7Hz trajectory generation, regardless of the terrain types, while generating 2 seconds long trajectory, which is reasonable as an online planner for quadrupedal robots.