Robust Motion Control of Robotic Systems with Environmental Interaction via Data-Driven Inversion of CPG

We propose CPG (central pattern generator) based robot control framework constructed based on: 1) the inverse model of the CPG that directly encodes resultant body motion of the robot in CPG parameters; and 2) the CPG parameter adaptation law that enforces robustness of the controller. These two components behaves as feedforward and feedback control for the CPG controlled robot, enabling us to achieve fast and robust generation of a desired body motion for robotic systems operated within complex environmental interaction. The inverse CPG model is constructed based on neural network along with autoencoder to efficiently deal with the dimension decrease from input to output of the model. Also, the CPG parameter adaptation is done with a concept of backpropagation, which is enabled by the adoption of smooth activation function for the inverse CPG model. For the development and verification of the proposed framework, simulation is conducted with two robotic systems, snake-like robot and pivotboard system.