Abstract The control of complex nonlinear robotic systems is often considered computationally intensive, typically relying on 32-bit or more powerful hardware. This work challenges that paradigm by presenting a novel implementation of a fully data-driven intelligent controller on a resource-constrained microcontroller for robotics. A ball-and-plate system coupled to a six degree-of-freedom Stewart–Gough parallel manipulator, a widely used system for evaluating nonlinear control strategies, was controlled using an artificial neural network (ANN) deployed on an Arduino UNO R3. The datasets were created through dedicated experiments, comprising over 5 h of operation, 1110 repetitions, and 243,000 data points which included key parameters such as target and measured position, elapsed time, servomotor angles, and program flags generated by a baseline PID controller. The ANN architecture featured six inputs, one hidden layer with 12 neurons, and six outputs, and was ported to the 8-bit Arduino UNO using the AIfES (Artificial Intelligence for Embedded Systems) framework. Comparative experiments against proportional-integral-derivative (PID) control demonstrated that, while maintaining comparable performance in nominal conditions, the ANN controller exhibited superior stability and robustness under nonlinear and extreme scenarios, especially under increased speeds that break the linearized PID assumptions, making it unstable. The ANN also demonstrated the ability to incorporate actuation saturation into the control model. These results demonstrate the superior robustness of intelligent controllers for nonlinear control, while highlighting the feasibility of deploying data-driven intelligent controllers on ultra-low-power embedded devices, broadening the application space of Tiny Machine Learning (TinyML) in real-time robotics.
Lüdtke et al. (Fri,) studied this question.