Distance-Controllable Long Jump of Quadruped Robot Based on Parameter Optimization Using Deep Reinforcement Learning

Quadruped robots interact with the ground with discrete foot points during locomotion, which makes them gain an advantage in obstacle crossing compared with the wheeled and tracked robots. Quadruped robots can jump from current position to one position a certain distance ahead to negotiate the obstacles between them, for example. However, current quadruped control strategies usually assume that the landing area is large enough, and thus jumping distance control of quadruped robots had not yet been studied sufficiently. This paper proposes a method for controlling the distance of quadruped robot jumps based on deep reinforcement learning (DRL). In the method, kinematic parameters in the control module are optimized to achieve the quadruped jumping tasks. Based on the understanding of the kinematics and dynamics of quadruped robot jumping, an initial jumping is realized by controlling the robot foot moving along a carefully designed parameterized trajectory. This initial trajectory is then used to train a set of jumping parameters using a deep reinforcement learning (DRL) algorithm. Through thousands of jumping trials in the Gazebo simulation environment, the optimal parameters were acquired. Our proposed method allows for accurate jumping within the 0.5 m to 0.8 m range. Additionally, the controller has been successfully implemented on a real quadruped robot.