Imitation learning from model predictive control for dynamic motions of quadruped

Abstract

Model Predictive Control(MPC), which computes control inputs to minimize a given cost function, has demonstrated high performance in various studies related to quadrupedal robots. However, the computation time needs to be shorter than the control period for real-time implementation of MPC. Therefore, applying high control frequency or long prediction horizons to MPC is limited. Research that utilizes imitation learning to replace MPC with an artificial neural network of comparable performance has been conducted to reduce computation time significantly. However, existing research has been applied to the walking motion of quadrupedal robots, where real-time control of MPC is possible without the need for high control frequency or long prediction horizons. This paper proposes imitation learning for MPC to implement various dynamic motions of a quadrupedal robot. In particular, by solving the problem of computation time through imitation learning, high performance resulting from the short control cycle and long prediction horizon of model predictive control can be achieved in real-time control. The motions implemented include jumping over obstacles while running with a bounding gait, jumping off a roof, etc. These motions are implemented with a single neural network.