# Evaluation of Different Controllers for Sensing-Based Movement Intention Estimation and Safe Tracking in a Simulated LSTM Network-Based Elbow Exoskeleton Robot

## TL;DR

This paper presents a new control system for elbow exoskeletons that uses muscle signals and an LSTM network to estimate and track movement intentions more accurately and safely.

## Contribution

The novel contribution is a hybrid LSTM–sliding mode control framework that improves real-time intention estimation and tracking in elbow exoskeletons.

## Key findings

- The LSTM network achieved an R2 of 0.965 and a 47% improvement over traditional methods.
- The sliding mode controller outperformed PID and impedance controllers with minimal tracking errors (average RMSE = 0.21 Nm).
- PID controllers showed significant errors and are unsuitable for direct myoelectric control.

## Abstract

Control of elbow exoskeletons using muscular signals, although promising for the rehabilitation of millions of patients, has not yet been widely commercialized due to challenges in real-time intention estimation and management of dynamic uncertainties. From a practical perspective, millions of patients with stroke, spinal cord injury, or neuromuscular disorders annually require active rehabilitation, and elbow exoskeletons with precise and safe motion intention tracking capabilities can restore functional independence, reduce muscle atrophy, and lower treatment costs. In this research, an intelligent control framework was developed for an elbow joint exoskeleton, designed with the aim of precise and safe real-time tracking of the user’s motion intention. The proposed framework consists of two main stages: (a) real-time estimation of desired joint angle (as a proxy for movement intention) from High-Density Surface Electromyography (HD-sEMG) signals using an LSTM network and (b) implementation and comparison of three PID, impedance, and sliding mode controllers. A public EMG dataset including signals from 12 healthy individuals in four isometric tasks (flexion, extension, pronation, supination) and three effort levels (10, 30, 50 percent MVC) is utilized. After comprehensive preprocessing (Butterworth filter, 50 Hz notch, removal of faulty channels) and extraction of 13 time-domain features with 99 percent overlapping windows, the LSTM network with optimal architecture (128 units, Dropout, batch normalization) is trained. The model attained an RMSE of 0.630 Nm, R2 of 0.965, and a Pearson correlation of 0.985 for the full dataset, indicating a 47% improvement in R2 relative to traditional statistical approaches, where EMG is converted to desired angle via joint stiffness. An assessment of 12 motion–effort combinations reveals that the sliding mode controller consistently surpassed the alternatives, achieving the minimal tracking errors (average RMSE = 0.21 Nm, R2 ≈ 0.96) and showing superior resilience across all tasks and effort levels. The impedance controller demonstrates superior performance in flexion/extension (average RMSE ≈ 0.22 Nm, R2 > 0.94) but experiences moderate deterioration in pronation/supination under increased loads, while the classical PID controller shows significant errors (RMSE reaching 17.24 Nm, negative R2 in multiple scenarios) and so it is inappropriate for direct myoelectric control. The proposed LSTM–sliding mode hybrid architecture shows exceptional accuracy, robustness, and transparency in real-time intention monitoring, demonstrating promising performance in offline simulation, with potential for real-time clinical applications pending hardware validation for advanced upper-limb exoskeletons in neurorehabilitation and assistive applications.