Bionic Multi-legged Robot Based on End-to-end Artificial Neural Network Control

This paper aims at a conceptual design of a lightweight prototype for autonomous planetary surface exploration. Considering the complex bumpy surface on planets, we design a novel 12-legged reshaping robot inspired by sea urchin structure, which holds the potential to fit unstructured terrains using simple motor skills. The prototype realizes the omnidirectional motion and has the features of no overturning and high fault tolerance. The autonomous locomotion policy is proposed based on a model-free end-to-end reinforcement learning algorithm with only proprioception, holding the feature of fast training and no prior knowledge. The robot with learned policy enables steady autonomous mobility and robust adaptation to the generalized terrains and external perturbation through the virtual simulation experiments in various unstructured environments. Finally, we organized prototype experiments in the laboratory, which validate the dynamic feasibility of the gait to directly deploy.