Robust Data-Driven Passivity-Based Control of Underactuated Systems via Neural Approximators and Bayesian Inference

Abstract

We synthesize controllers for underactuated robotic systems using data-driven approaches. Inspired by techniques from classical passivity theory, the control law is parametrized by the gradient of an energy-like (Lyapunov) function, which is represented by a neural network. With the control task encoded as the objective of the optimization, we systematically identify the optimal neural net parameters using gradient-based techniques. The proposed method is validated on the cart-pole swing-up task, both in simulation and on a real system. Additionally, we address questions about controller’s robustness against model uncertainties and measurement noise, using a Bayesian approach to infer a probability distribution over the parameters of the controller. The proposed robustness improvement technique is demonstrated on the simple pendulum system.