A novel adaptive dynamic optimal balance control method for wheel-legged robot

Abstract

The wheel-legged robots combine efficient and fast wheeled locomotion with the terrain-adaptive legged locomotion. Inspired by reinforcement learning and adaptive dynamic programming, a novel dynamic optimal balance control method is proposed for wheel-legged robots on uneven terrains. First, the virtual leg length is solved according to the kinematics model of the five-link closed-chain mechanism. In addition, a simplified wheel-legged spring-loaded inverted pendulum model is established to determine the linear state-space expression of the floating-base, virtual leg, and driving wheel. Second, a fast iterative algorithm built upon adaptive dynamic programming and optimal gain matrix is introduced. Using the initial gain matrix and an initial state vector, the online policy iteration learns the initial state data set generated by external disturbances, and the steps of policy evaluation and policy improvement are repeatedly implemented by Kleinman's algorithm. Subsequently, the virtual support force is controlled by the composite control framework for the length of the virtual leg with spring-damping characteristics and roll angle. The input torque for each hip joint is calculated using the virtual model control mapping technology. Finally, the robustness and adaptability of the proposed framework are verified through simulations. This paper presents a novel control method for the future application of wheel-legged robot in complex scenarios.