Torque tracking position control of DLR-HIT II robotic hand using a real-time physics-informed neural network

Abstract

This paper presents a novel approach for controlling the DLR-HIT II robotic hand by leveraging physics-informed neural networks (PINNs) for torque and position control. This method eliminates the need for additional control inputs or external controllers, achieving high precision and simplified dynamics, which is validated through extensive simulations that closely replicate experimental conditions, demonstrating the system's ability to handle external disturbances and maintain accurate trajectory tracking. The strategy only requires time and joint position data as inputs, allowing the network to compute velocity and acceleration internally. Time normalization enhances the model's ability to generalize across different time scales and ensures stable training. The method demonstrates strong generalization from a limited training set and successfully performs across diverse trajectory types. This simplification significantly reduces computational complexity and facilitates real-time control in advanced robotic applications.