Data-Driven Design of Energy-Shaping Controllers for Swing-Up Control of Underactuated Robots

We propose a novel data-driven procedure to train a neural network for the swing-up control of underactuated robotic systems. Our approach is inspired by several recent developments ranging from nonlinear control theory to machine learning. We embed a neural network indirectly into the equations of motion of the robotic manipulator as its control input. Using familiar results from passivity-based and energy-shaping control literature, this control function is determined by the appropriate gradients of a neural network, acting as an energy-like (Lyapunov) function. We encode the task of swinging-up robotic systems through the use of transverse coordinates and goal sets; which drastically accelerates the rate of learning by providing a concise target for the neural network. We demonstrate the efficacy of the algorithm with both numerical simulations and experiments.