Real time gait planning for a quadruped robot using imitation learning

Abstract

This thesis presents a method to obtain a gait planning that is suitable for quadruped robots and executable in real-time. The proposed idea is to use a neural network to obtain the gait sequence. The neural network is trained through Imitation Learning to learn the optimal contact sequence by Monte Carlo Tree Search (MCTS) algorithm. Specifically, the Supervised Learning approach is adopted in the initial stage, and subsequently, the results are improved by implementing DAGGer algorithm. The replacement of the MCTS algorithm with a neural network causes a big gain in terms of computational time. With this reduced calculation time, the gait sequence can be calculated in real-time.