A Feature Extraction Pipeline for Enhancing Lightweight Neural Networks in sEMG-based Joint Torque Estimation This work was funded by the German Federal Ministry of Education and Research (BMBF) under Grant No. 01IW24008.
Kartik Chari 1 2 , Raid Dokhan 2 , Anas Homsi 2 , Niklas Kueper 1 2 , Elsa Andrea Kirchner 1 2 · 2026
Read A Feature Extraction Pipeline for Enhancing Lightweight Neural Networks in sEMG-based Joint Torque Estimation This work was funded by the German Federal Ministry of Education and Research (BMBF) under Grant No. 01IW24008. by Kartik Chari 1 2 , Raid Dokhan 2 , Anas Homsi 2 , Niklas Kueper 1 2 , Elsa Andrea Kirchner 1 2 (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-16712.txt.
Introduction
Kartik Chari, Raid Dokhan, Anas Homsi, Niklas Kueper, and Elsa Kirchner aid rehab with smart sleeves. Exoskeletons can lift and guide weak arms. To help well, the frame must guess the force at each joint. Their 2026 preprint draws that guess from skin signals. Eight small pads read the hum of arm muscles.
Light neural nets turn those hums into force calls. Such nets must run fast on small gear. The team built a clean path from raw hum to force. Their path picks sharp traits from the signal first. Good force calls let frames help, not fight, the user. This could aid stroke and injury rehab in clinics.
Catalog record
| CREATOR | Kartik Chari 1 2 , Raid Dokhan 2 , Anas Homsi 2 , Niklas Kueper 1 2 , Elsa Andrea Kirchner 1 2 |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 31 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2601-16712
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
A Feature Extraction Pipeline for Enhancing Lightweight Neural Networks in sEMG-based Joint Torque Estimation This work was funded by the German Federal Ministry of Education and Research (BMBF) under Grant No. 01IW24008. Kartik Chari 1 2 , Raid Dokhan 2 , Anas Homsi 2 , Niklas Kueper 1 2 , Elsa Andrea Kirchner 1 2 Abstract Robot-assisted rehabilitation offers an effective approach, wherein exoskeletons adapt to users’ needs and provide personalized assistance. However, to deliver such assistance, accurate prediction of the user’s joint torques is essential. In this work, we propose a feature extraction pipeline using 8-channel surface electromyography (sEMG) signals to predict elbow and shoulder joint torques. For preliminary evaluation, this pipeline was integrated into two neural network models: the Multilayer Perceptron (MLP) and the Temporal Convolutional Network (TCN). Data were collected from a single subject performing elbow and shoulder movements under three load conditions (0 kg, 1.10 kg, and 1.85 kg) using three motion-capture cameras. Reference torques were estimated from center-of-mass kinematics under the assumption of static equilibrium. Our offline analyses showed that, with our feature extraction pipeline, MLP model achieved mean RMSE of 0.963 N m, 1.403 N m, and 1.434 N m (over five seeds) for elbow, front-shoulder, and side-shoulder joints, respectively, which were comparable to the TCN performance. These results demonstrate that the proposed feature extraction pipeline enables a simple MLP to achieve performance comparable to that of a network designed explicitly for temporal dependencies. This finding is particularly relevant for applications with limited training data, a common scenario patient care. I Introduction True human-robot collaboration requires co-adaptation, where both humans and robots adapt to each other’s behaviors. Traditionally, assistive robots contributed with physical strength while humans provided cognitive skills [ 24 , 19 , 9 ] . Today, robots are increasingly expected to simulate human behaviors, recognize actions, and anticipate intentions [ 13 ] , but human variability and physiological unpredictability remain major challenges. In rehabilitation, accurate tracking and prediction of joint movements are critical. Some studies measured joint torques directly using sensors and mechanical devices [ 11 ] , while others relied on musculoskeletal models such as the Hill-type model [ 18 , 8 ] . However, as noted in [ 9 , 21 ] , such models are difficult to construct, since they require numerous parameters to be measured using specialized equipment. Surface electromyography (sEMG)-driven approaches simplify computation by combining muscle models with body dynamics [ 27 ] . Building on this, numerical methods were developed to associate predefined hand motions with corresponding forearm EMG signals [ 17 ] . However, these methods struggled to capture nonlinear muscle dynamics. As a result, machine learning (ML) approaches gained attention for joint torque estimation. For instance, [ 21 ] used a backpropagation neural network (BPNN) to predict elbow torques from EMG whereas [ 33 ] estimated ankle torques using an EMG-driven neuromusculoskeletal (NMS) model combined with an artificial neural network (ANN). More recent works applied recurrent convolutional networks (RCNN) to predict knee torques [ 26 , 10 ] , while EMG-driven approaches with NMS models have been explored for the upper limb but they require precise parameter estimation, limiting their practicality [ 29 ] . To address this, EMG-based feature extraction methods have emerged. Some studies have used statistical methods and manually extracted features to capture muscle dynamics and improve signal interpretation [ 15 , 22 ] . Autoencoders have been used to extract informative features [ 9 ] , convolutional neural networks (CNN) and temporal convolutional networks (TCN) have demonstrated superior performance in modeling temporal dependencies without explicit feature engineering [ 3 , 16 ] . Nevertheless, these methods often require large datasets, complex pipelines, and computationally expensive architectures, which are impractical in rehabilitation contexts due to patient fatigue and limited data. Motivated by these challenges, we propose a dedicated EMG feature extraction pipeline for estimating joint torques. Our pipeline enabled accurate predictions even with a simple multilayer perceptron (MLP) network. To evaluate whether the extracted features effectively capture temporal dynamics, we compared their performance with that of TCN, a model explicitly designed for sequence modeling using raw, preprocessed time points [ 31 ] . II EXPERIMENTAL DESIGN This section provides details about the participant, the experimental setup, data acquisition methods, and the experimental procedure. II-A Participant One healthy male participant (24 years old, student) voluntarily took part in this study. A few days before the experiment, he was familiarized with the setup and procedure. He provided written informed consent, including permission for recording and use of his images and videos. The participant was informed of his right to stop the experiment at any time without consequences. The study was approved by the local ethics committee and conducted in accordance with the local legislation and institutional requirements. II-B Experimental Setup Figure 1a shows the setup used for data collection. The participant sat on a comfortable chair holding an object of mass 0\text{\,}$\mathrm{k}\mathrm{g}$ , 1.10\text{\,}$\mathrm{k}\mathrm{g}$ , or 1.85\text{\,}$\mathrm{k}\mathrm{g}$ . The objects used as weights in the experiment are shown in Figure 1b . Three Qualisys Oqus 300 cameras tracked the participant’s joint trajectories with the help of six passive markers at a sampling rate of 500\text{\,}$\mathrm{H}\mathrm{z}$ . These markers were placed on the shoulders, elbows, and wrists of both arms, and the cameras tracked them from multiple viewpoints to obtain precise 3D positions of each marker. (a) Participant performing a task (b) Objects used as weights Fig. 1: Experimental setup for data collection. (a) Participant prepared with EMG electrodes lifting an object; two of the three Qualisys motion-capture cameras are visible. (b) Weighted objects used in the experiment with measured masses of 1.10\text{\,}$\mathrm{k}\mathrm{g}$ and 1.85\text{\,}$\mathrm{k}\mathrm{g}$ . The 0\text{\,}$\mathrm{k}\mathrm{g}$ weight is not shown. Furthermore, the ANT mini eego amplifier was used to record bipolar surface EMG data at a sampling rate of 500\text{\,}$\mathrm{H}\mathrm{z}$ with the help of eego SDK for Python. Eight EMG channels were recorded, and the electrodes were placed on the belly of each muscle after preparing the skin with 70% v/v isopropyl alcohol, in accordance with the SENIAM guidelines [ 7 ] . The electrodes measured the following muscles of the right arm: deltoid pars clavicularis, deltoid pars acromialis, deltoid pars spinalis, biceps brachii, triceps brachii caput laterale, triceps brachii caput mediale, brachialis, and pectoralis major. II-C Experimental Procedure Fig. 2: Flowchart of the experimental procedure. In each repetition, the participant performed the movements with weights chosen in ascending order. Movement 1 denotes grasping, and Movement 2 denotes complex movement. The aim of this experiment was to collect EMG and motion data from a healthy participant under varying levels of muscle activation in the shoulder and elbow. These data were intended as inputs to a machine learning regressor model for predicting human joint torques from their EMG signals, with motion data providing the reference torque values for supervised learning. To elicit these varying muscle activation levels, we selected three different weights ( 0\text{\,}$\mathrm{k}\mathrm{g}$ , 1.10\text{\,}$\mathrm{k}\mathrm{g}$ , and 1.85\text{\,}$\mathrm{k}\mathrm{g}$ ) and two movement types, namely, grasping and complex movement. The grasping movement consisted of moving the arm from rest to full extension in the sagittal plane and back, mimicking the action of grasping an object on a table. Similarly, the complex movement combined