[Paper] Lightweight Test-Time Adaptation for EMG-Based Gesture Recognition

Source: arXiv - 2601.04181v1

Overview

Surface electromyography (EMG)‑based gesture recognition is the backbone of many next‑generation wearables and prosthetic controllers. However, the signal drifts that occur when electrodes shift, muscles fatigue, or users change posture cause a sharp drop in accuracy once a model leaves the lab session it was trained on. This paper presents a lightweight test‑time adaptation (TTA) framework that lets a compact Temporal Convolutional Network (TCN) stay accurate across sessions without the heavy data‑collection or compute budgets typical of existing solutions.

Key Contributions

• Causal Adaptive Batch Normalization (AdaBN): A real‑time, on‑device method that continuously aligns the statistics of incoming EMG streams to the model’s learned distribution.
• GMM‑Based Alignment with Experience Replay: A probabilistic alignment that updates a Gaussian Mixture Model of feature embeddings while replaying a tiny buffer of past data to avoid catastrophic forgetting.
• Meta‑Learning Calibration: A few‑shot meta‑training scheme that enables the model to adapt to a new session with just one or two labeled gestures, cutting calibration data by an order of magnitude.
• Comprehensive Evaluation on NinaPro DB6: Demonstrates that the proposed TTA strategies close the inter‑session accuracy gap by up to ~20 % while adding only a few milliseconds of latency and negligible memory overhead.
• Deployment‑Ready Design: All components are compatible with low‑power microcontrollers, making “plug‑and‑play” myoelectric control feasible for everyday prosthetic use.

Methodology

1. Backbone Model – Temporal Convolutional Network (TCN)

   • Chosen for its ability to capture temporal dependencies in EMG signals with a small parameter count.
   • Operates causally, meaning predictions rely only on past samples—crucial for real‑time control.

2. Test‑Time Adaptation Strategies

   • Adaptive Batch Normalization: During inference, the running mean/variance of each BN layer are updated with a moving average of the current batch (or single sample) statistics. Because the updates are causal, the system never looks ahead, preserving low latency.
   • GMM Alignment + Experience Replay: Feature vectors from the TCN are modeled as a mixture of Gaussians. At test time, the model refines the GMM parameters using incoming data while periodically replaying a fixed‑size buffer (≈ 50 samples) of previously seen embeddings to keep the decision boundaries stable.
   • Meta‑Learning (MAML‑style) Calibration: The TCN is meta‑trained across many simulated session shifts. At deployment, a gradient step (or two) on a handful of labeled gestures instantly re‑tunes the network for the new session.

3. Training & Evaluation Pipeline

   • Pre‑train the TCN on a large portion of the NinaPro DB6 multi‑session dataset.
   • Freeze the backbone and apply each TTA method independently during the test sessions.
   • Measure classification accuracy, latency, and memory footprint across 10‑fold cross‑session splits.

Results & Findings

• Stability: The experience‑replay variant showed the least variance across sessions, confirming its ability to prevent forgetting.
• Data Efficiency: Meta‑learning achieved >70 % accuracy with only one or two labeled gestures, whereas prior state‑of‑the‑art calibration required dozens of samples.
• Resource Footprint: All three TTA approaches added less than 5 ms of processing time on a Cortex‑M4 microcontroller and required under 1 KB of extra RAM, well within the constraints of typical prosthetic hardware.

Practical Implications

• Plug‑and‑Play Prosthetics: Users can put on a myoelectric device, perform a couple of quick calibration gestures, and immediately benefit from robust control without a lengthy training session.
• Energy‑Efficient Wearables: The low‑overhead adaptation fits on battery‑powered edge devices, extending operational time compared to heavyweight domain‑adaptation pipelines.
• Rapid Deployment for New Users: Manufacturers can ship a single “universal” model and rely on on‑device TTA to personalize performance, reducing the need for per‑user model training in the cloud.
• Generalizable Framework: The causal AdaBN and GMM‑replay ideas are not EMG‑specific; they can be transplanted to other streaming sensor domains (e.g., EEG, inertial measurement units) where distribution shift is a pain point.

Limitations & Future Work

• Limited to Single‑Channel Adaptation: The current TTA updates only the BN statistics and feature‑space GMM; deeper architectural changes (e.g., weight adaptation) were not explored.
• Small Replay Buffer: While memory‑friendly, a 50‑sample buffer may be insufficient for highly non‑stationary users; adaptive buffer sizing could improve robustness.
• Evaluation on One Dataset: Results are reported on NinaPro DB6; broader validation on other EMG corpora and real‑world prosthetic trials is needed.
• Meta‑Learning Overhead: The meta‑training phase is computationally intensive and must be performed offline; future work could investigate online meta‑learning to further reduce pre‑deployment costs.

Bottom line: This paper shows that smart, lightweight test‑time adaptation can bridge the gap between lab‑grade EMG models and reliable, everyday prosthetic control—opening the door for truly plug‑and‑play myoelectric devices that adapt on the fly with minimal compute and data.

Authors

• Nia Touko
• Matthew O A Ellis
• Cristiano Capone
• Alessio Burrello
• Elisa Donati
• Luca Manneschi

Paper Information

• arXiv ID: 2601.04181v1
• Categories: cs.LG, cs.HC
• Published: January 7, 2026
• PDF: Download PDF