Handling the Transition in the Locomotion of an Articulated Quadruped Robot by Adaptive CPG

摘要

A control strategy based on the theory of Central Pattern Generators (CPG) implemented through recurrent neural network (RNN) for the development of trotting and walking locomotion of an articulated quadruped robot is presented in this paper. The control strategy was enriched with fuzzy logic to ensure the stability and balance of the robot in transitions among locomotion modes. The control strategy was simulated and then implemented in a prototype. The results are very promising. A quadruped robot with three joints per leg, name R4A3, was built for this purpose. The R4A3 learnt to walk by itself; for its locomotion, it is not necessary to develop the kinematic equations. A novel Nonlinear First Order Differential Equation System was developed to simulate the CPG as oscillators. The control strategy is distributed and autonomous, operating through three levels of control. Each level determines the degree of complexity of the action concerning the consequence of how it affects the entire robot. Level 1 aims to those repetitive continuous movements occurring on each joint. These movements are action primitives. Level 2 governs repetitive movements occurring on each leg. It involves the basic actions that identify the type of locomotion. Level 3 considers the recognition of the motion of the R4A3 determining its stability and balance, planning transitions between types of locomotion and avoiding overturning. With the appropriated parameterization the same Differential Equation System is used in each CPG on each level reducing the mathematical complexity to reproduce the quadruped movement resulting in a low computational cost suitable for hardware implementation.