Martin Colot

This demo shows real-time control of a robotic hand (Leap Hand) from 8 sEMG sensors around the forearm using regression of the finger joint angles. The recognition model is first trained from 10 minutes of EMG data from the test subject, labeled using a Leap Motion camera . Figure 1. A. The MindRove 8-channel EMG armband system to record the sEMG signals. B. The Leap Motion Controller to estimate the hand poses. C. The Leap Hand robotic hand that represents the final physical effector. D. Synchronized sEMG signals with three representative finger joint angles. Watch the Demo Figure 2. Visualization of the true (black) versus predicted (blue) finger joint angles over time.

This early demo shows my latest results of hand gesture recognition from 8 EMG signals and a single gyroscope . I’m using the MindRove EMG armband for data acquisition. MindRove EMG armband used in the experiment Before recording the video, I collected about 10 minutes of labeled training data using a LeapMotion camera. I then trained a machine learning model (details to be published) and ran inference without relocating the sensors. Live demo of real-time gesture prediction Prediction example (Black = Real Angle, Blue = Predicted, Red = Error)

In this paper, we introduce the novel LDA-KM-DA algorithm — a non-conservative unsupervised domain adaptation method designed to improve EMG-based hand gesture recognition across different subjects. The non-conservative nature of the method means the model is adapted using only samples from the test domain and initial pseudo-labels from source domains . This results in higher accuracy than comparable state-of-the-art approaches. ▶️ See it in action: Live Demo Illustration of LDA-KM-DA 📄 Read More Here: Colot, M. , Simar, C., Cebolla Alvarez, A. M., & Bontempi, G. Linear Non-Conservative Unsupervised Domain Adaptation for Cross-Subject EMG Gesture Recognition . Available at SSRN: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5099691 FREE ACESS UNTIL SEPTEMBER 03 2025 : https://www.sciencedirect.com/science/article/pii/S1746809425007943?dgcid=author

In this video, you can see a demonstration of my algorithm LDA-KM-DA , which performs non-conservative unsupervised domain adaptation to adapt an EMG-based hand gesture recognition model from a set of training subjects to a new test user. Initially, the model performs poorly. But as it processes more EMG data, it quickly learns and improves — and by the end, gesture recognition is nearly perfect . 🔗 Learn more: Full Article

This paper is the result of the work done during the Trail Summer Workshop 2024 in Lisbon. We show how to use machine learning to define the domain of operation of an autonomous car in terms of weather parameters from camera images by predicting complete probability distributions instead of single values. 📄 Read More Here : Halin, A. , Piérard, S., Vandeghen, R., Gérin, B., Zanella, M., Colot, M., ... & Van Droogenbroeck, M. Physically Interpretable Probabilistic Domain Characterization . In Proceedings of the Asian Conference on Computer Vision (pp. 15–33), 2024: https://openaccess.thecvf.com/content/ACCV2024W/AWSS/html/Halin_Physically_Interpretable_Probabilistic_Domain_Characterization_ACCVW_2024_paper.html

EMG Dataset and Research In this paper, Cédric and I present a novel dataset for hand gesture recognition based on EMG signals from the forearms. The dataset includes EMG recordings from both arms, along with motion capture data from 14 participants performing various tasks in virtual reality. It can be used for gesture classification in guided scenarios or for regression of finger joint angles during unrestricted interactions with the virtual environment. We also evaluate intra-subject classification models using different feature extraction methods. In particular, we investigate how physiologically informed feature extraction can improve model accuracy by separating the EMG signals into Move Command and Hold Command phases. EMG Feature Extraction Illustration Read More Here: Simar, C. , Colot, M. , Cebolla, A. M., Petieau, M., Cheron, G., & Bontempi, G. Machine learn...

I presented this work at the workshop Adapting to Change: Reliable Multimodal Learning Across Domains during ECML-PKDD 2023. As the basis of my PhD thesis, this paper highlights the cross-subject issue in hand gesture recognition from EMG signals of the forearm. Due to electrode placement shifts, fatigue, and sweat, the class-conditional distributions of the samples change across users. This causes a drop in accuracy in cross-subject configurations, where a classifier is tested on a subject unseen during training. We show how to reduce the accuracy gap using Subspace Alignment , a simple unsupervised domain adaptation method. Read More Here : Colot, M. , Simar, C., Petieau, M., Cebolla Alvarez, A. M., Cheron, G., & Bontempi, G. EMG subspace alignment and visualization for cross-subject hand gesture classification . In Joint European Conference on Machine Learning and Knowledge Discovery in Databases (pp. 416–423). Cham: Springer Na...

🎉 Welcome to my website! I’m delighted to share with you my research work on exciting topics such as machine learning applied to neurophysiology, intelligent robotic prosthetics, and transfer learning. Here you'll find my latest scientific publications, interactive demos, images, videos… and some personal reflections along the way! 💬 Have a question, an idea, or a collaboration in mind? 📩 Contact me at martin.colot@ulb.be Also, discover more work from my lab :  Machine Learning Group