Neural Human Intent Estimator for an Adaptive Robotic Nursing Assistant

Abstract

Estimation of human intent during interaction with a robot is important for maintaining safety, predictability, and performance. This paper proposes a neural, model-free, online human intent estimator to guide our Adaptive Robotic Nursing Assistant (ARNA) robot. ARNA is a service mobile manipulator designed to assist nurses and healthcare workers with patient sitting and walking tasks in hospital environments. The proposed Human Intent Estimator (HIE) is implemented as two efficient one-layer neural networks (NN) that generate reference trajectories based on the user torque inputs from the robot’s handlebar. These trajectories are sent to an NN-based neuroadaptive controller (NAC) in the inner loop to generate the necessary wheel torques to follow human-guided trajectories. The proposed NN weight adaptation laws for the HIE-NAC are shown to be stable as long as the trajectories generated by the human intent dynamics are smooth and bounded, which requires the user not to change his/her intention abruptly. We tested the proposed intent estimator and controller in three different human-robot interaction experiments with 10 participants. A linear mixed-effect model was used to test the difference between the HIE-NAC scheme and a conventional admittance controller. The results show significant performance improvements by reducing jerk and tracking velocity errors during operation.