Approximate Neural Network-based Nonlinear Model Predictive Control of Soft Continuum Robots

This paper employs a deep learning approach to enhance the task space control of soft continuum robots. We built an approximate data-driven dynamics model of a soft robot using sampled Cartesian positions of the robot's tip and its actuator tensions. We then incorporated this surrogate model into a Model Predictive Control (MPC) control scheme, enabling nonlinear control for task space trajectory tracking without relying on an exact analytical dynamics model of the robot or extensive computations. By involving constraints into the MPC, we addressed the robot's workspace and actuation limits. The numerical results of the simulation experiment show that deep learning dynamic models can improve robotic control. This leads to accurate trajectory tracking and suggests that deep learning could be used more in robot system control, especially for realtime control applications.