Design of Minimal Model-Free Control Structure for Fast Trajectory Tracking of Robotic Arms

Abstract

This paper designs a minimal neural network (NN)-based model-free control structure for the fast, accurate trajectory tracking of robotic arms, crucial for large movements, velocities, and accelerations. Trajectory tracking requires an accurate dynamic model or aggressive feedback. However, such models are hard to obtain due to nonlinearities and uncertainties, especially in low-cost, 3D-printed robotic arms. A recently proposed model-free architecture has used an NN for the dynamic compensation of a proportional derivative controller, but the minimal requirements and optimal conditions remain unclear, leading to overly complex architectures. This study aims to identify these requirements and design a minimal NN-based model-free control structure for trajectory tracking. Two architectures are compared, one NN per joint (INN) and one global NN (GNN), each tested on two serial robotic arms in simulations and real scenarios. The results show that the architecture reduces tracking errors (RMSE < 2°). The INN is accurate for decoupled joint dynamics and requires fewer training data than the GNN. A table summarizes the design process. Future works will apply this control structure to low-cost robotic arms and micro-movements.