# Integrating contrastive cross-modal attention and stacked GRU for hand function rehabilitation robot control

## TL;DR

This paper introduces a new model for hand rehabilitation robots that improves control accuracy by combining attention mechanisms and recurrent neural networks.

## Contribution

The novel C-GAP model integrates cross-modal attention and stacked GRUs for improved hand movement control in rehabilitation robots.

## Key findings

- The C-GAP model effectively captures dynamic temporal features of hand movements using stacked GRUs.
- The model performs stably on both healthy and stroke patient datasets, adapting to different hand function states.
- It provides a complete control scheme for rehabilitation robots with adaptive PID optimization.

## Abstract

With the intensification of population aging and the increasing incidence of neurological diseases, the demand for precise and intelligent control technology in hand rehabilitation robots has become more urgent. Traditional control methods struggle to effectively capture the dynamic temporal features of hand movements, especially in scenarios where there are modal differences between hand function data of healthy individuals and stroke patients, leading to insufficient control accuracy and poor generalization. This paper focuses on hand rehabilitation robot control technology and proposes the C-GAP model. By designing a cross-modal attention mechanism, the model realizes feature collaboration of multi-source data such as electromyography (EMG), force, and joint angles. It relies on Stacked Gated Recurrent Units to accurately extract the temporal dynamic patterns of typical hand functional movements, such as grasping, pinching, and wrist rotation. In combination with an adaptive PID controller, the model optimizes force-controlled trajectories in rehabilitation training, forming a complete control scheme tailored to hand rehabilitation scenarios. Experimental validation shows that the model performs stably in classifying typical hand functional movements and dynamic control on the Ninapro DB5 (healthy hand function multimodal data) and MUSED-I (stroke patient hand function unimodal data) datasets, effectively adapting to rehabilitation training needs under different hand function states. The research provides technical support for the precise perception and control of sequential movements in hand rehabilitation robots, contributing to enhancing the specificity and safety of rehabilitation training. It has practical significance for promoting the application of recurrent neural networks in the field of rehabilitation robot control.