Control of robot manipulators with uncertain closed architecture using neural networks

This paper presents a novel neural network-based control approach designed for industrial robot manipulators characterized by uncertain closed architectures and unknown dynamics. Industrial and commercial robot manipulators typically employ closed control architectures, which limit the ability to make modifications or comprehend the inner control processes. Users are generally restricted to providing joint position or velocity commands for controlling the manipulator. Furthermore, the integration of these robots with external sensors for modern applications poses challenges to system stability. Our proposed solution utilizes neural networks to approximate the robot's dynamic model and low-level controller. The proposed controller is introduced as an outer (external feedback) loop, ensuring independence from the inner controller configuration. This outer loop leverages external sensor data and the desired trajectory to calculate commands for joint velocities. Consequently, this approach offers greater design flexibility for modern control applications. Unlike previous studies, our work introduces novelty through unconstrained control actions, avoiding the need for inner controller configuration and control gain structure. To validate our method, we conducted experiments using two industrial manipulators, namely the UR5e and UR10e, and the results clearly demonstrate the superior performance and industrial applicability of the framework we have developed.