Sliding Mode Control and Observation for Upper Limb Robotics Rehabilitation

Abstract

Stroke is a leading cause of upper limb motor impairments, which often require rehabilitation. General purpose robotic manipulators enable effective therapy through exercises such as Center Out Point‐to‐Point (CO‐PTP) and Teach‐and‐Track (TT), which respectively demand accurate estimation and rejection of the interaction force exerted by the patient on a handle mounted at the robot end‐effector (EE). Sliding Mode Control (SMC) represents a valid solution to serve these purposes due to the matched nature of the torque disturbances derived from the interaction force. In this paper, we describe and evaluate different model‐based SMC observation and control strategies during rehabilitation exercises. Additionally, we propose a novel observation scheme that combines Deep Neural Networks (DNNs) with Integral Sliding Mode (ISM), and compare its performance with the model‐based approaches. All methods are validated through experimental tests using a Franka Emika Panda robot, focusing on passive TT execution and force estimation in CO‐PTP scenarios.