💻 computer science

GEGLU-Transformer for IMU-to-EMG Estimation with Few-Shot Adaptation

This paper proposes a GEGLU-Transformer framework that reconstructs continuous muscle activation envelopes from inertial measurement unit (IMU) data with rapid few-shot adaptation, offering a robust and scalable alternative to direct surface electromyography (EMG) sensing for wearable robotic control.

Original authors: Miroljub Mihailovic, Luca Tonin, Stefano Tortora, Emanuele Menegatti

Published 2026-04-29

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Miroljub Mihailovic, Luca Tonin, Stefano Tortora, Emanuele Menegatti

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot to walk like a human. To do this safely and naturally, the robot needs to know exactly when and how hard your muscles are working. Usually, we use special sticky sensors called EMG electrodes to "listen" to the electrical signals of your muscles. But here's the problem: these sensors are like sensitive microphones at a rock concert. If you move your arm slightly, sweat a little, or if the sticker shifts, the sound gets distorted. They are also very "picky," meaning a sensor setup that works perfectly for you might be useless for your friend.

This paper proposes a clever workaround: Instead of listening to the muscles directly, let's listen to the movement.

The Core Idea: The "Body Translator"

The researchers built a smart computer program (an AI) that acts like a translator.

The Input: It takes data from IMUs (Inertial Measurement Units). Think of these as the tiny motion sensors inside your smartphone or smartwatch. They are tough, cheap, and don't care about sweat or shifting stickers. They just measure how your legs are moving.
The Output: The AI tries to guess what your muscles would be doing based on those movements. It reconstructs the "muscle activation envelope"—basically, a smooth curve showing how hard each muscle is working over time.

The Problem: Everyone Moves Differently

The tricky part is that every human is unique. Your walking style, muscle strength, and bone structure are different from mine.

The Old Way: Most AI models are trained on a huge group of people and then just "frozen." When you put them on a new person, they often stumble because they haven't learned that specific person's quirks. It's like trying to speak to a foreigner using a dictionary that only has generic phrases; it works okay, but not perfectly.
The New Way: This paper introduces a system that can learn on the fly. It's like having a translator who can instantly pick up your accent after hearing just a few sentences.

The Secret Sauce: The "GEGLU-Transformer"

The team built a new type of AI brain called a GEGLU-Transformer.

The Transformer: Imagine a librarian who can read a whole book and instantly understand how the beginning connects to the end, even if there are hundreds of pages in between. This helps the AI understand long-term patterns in your walking, not just the immediate step.
The GEGLU (Gated Linear Unit): This is a special "gatekeeper" inside the AI. Think of it as a bouncer at a club who decides which information is important and which should be ignored. It helps the AI focus on the right signals and ignore the noise, making it much better at guessing muscle activity for different people.

The Magic Trick: "Few-Shot" Adaptation

This is the most exciting part. The researchers tested their system on a new person they had never seen before.

The Cold Start: First, they let the AI guess using only what it learned from everyone else. It was decent (about 70% accurate in matching the muscle patterns).
The "Few-Shot" Adjustment: Then, they showed the AI just 0.5% of that new person's walking data. To put that in perspective: if a person takes 2,400 steps in a test, the AI only needed to look at about 12 steps to learn how to adjust.
The Result: After seeing those tiny 12 steps, the AI's accuracy jumped significantly. It learned the person's specific style almost instantly.

What They Found

Better than the Old Guard: Their new AI (GEGLU-Transformer) was consistently better at guessing muscle activity than older, standard AI models (like LSTMs), even before it got a chance to learn the new person.
Fast Learner: The "0.5% data" trick worked wonders. The AI didn't need hours of calibration; it needed a few seconds of walking data to become personalized.
Robustness: It worked well whether the person was walking on flat ground, climbing stairs, or walking up a ramp.

The Bottom Line

The paper claims that by using this new "smart translator" architecture, we can estimate muscle activity using simple, rugged motion sensors instead of fussy muscle sensors. Crucially, the system can adapt to a new user almost instantly with very little data. This makes the idea of wearable robots (like exoskeletons for walking assistance) much more practical for real-world use, where you can't spend hours calibrating sensors for every single person.

1. Problem Statement

The core challenge addressed is the reliable estimation of neuromuscular activation (EMG) from wearable Inertial Measurement Units (IMUs) for use in adaptive wearable robotics (e.g., exoskeletons, prostheses).

Limitations of Direct EMG: While EMG provides direct muscle activation data, it is difficult to deploy outside laboratories due to signal non-stationarity, sensitivity to electrode placement, motion artifacts, and strong inter-subject variability.
Limitations of Current IMU-to-EMG Models: Existing learning-based approaches typically rely on static models trained offline on multi-subject datasets. These models often fail to generalize to unseen users because they do not account for individual neuromuscular strategies. Furthermore, they lack efficient mechanisms for rapid, subject-specific personalization (few-shot adaptation) in real-world scenarios involving diverse locomotion tasks (stairs, slopes, etc.).

2. Methodology

The authors propose an adaptive learning framework consisting of three main components:

A. Signal Processing

Data Source: Synchronized IMU (200 Hz) and EMG (1000 Hz) data from 22 healthy adults performing various tasks (level walking, stairs, ramps, treadmill).
Preprocessing:
- EMG: Band-pass filtered (20–450 Hz), full-wave rectified, and low-pass filtered (8 Hz) to extract smooth activation envelopes.
- Alignment: IMU and EMG signals are resampled to 200 Hz and aligned.
- Segmentation: Data is segmented into gait cycles (0–100%) and time-normalized to 101 samples.
- Normalization: Signals are scaled to [0, 1] and median-filtered to remove artifacts.

B. GEGLU-Transformer Architecture

The model replaces standard recurrent architectures (like LSTM) with a Transformer-based encoder designed for cross-subject generalization.

Convolutional Front-End: A 1D Convolutional layer (kernel size 5) projects 24 inertial features into a 256-dimensional latent space to capture local motion patterns.
Self-Attention Encoder: Four stacked Transformer encoder layers with multi-head self-attention (8 heads) capture long-range temporal dependencies.
GEGLU Feed-Forward: Instead of standard feed-forward networks, the model uses Gaussian Error Gated Linear Units (GEGLU).
- Mechanism: $FF_{GEGLU}(x) = W_o(GELU(W_1x) \odot (W_2x))$ .
- Benefit: The gating mechanism modulates channel-wise activations, improving feature selection and gradient flow compared to sigmoid-based gating, which is crucial for regression tasks.
Output: A linear projection maps the latent representation to the target EMG envelopes for multiple muscles.

C. Few-Shot Adaptation

To handle unseen subjects, the framework employs a lightweight adaptation strategy:

Protocol: A small subset of the test subject's data (e.g., 0.5% to 10% of gait cycles) is used for calibration.
Process: The pre-trained model parameters are fine-tuned using gradient descent (AdamW) on this small adaptation set.
Constraints: All parameters are updated (no layer freezing) to allow full personalization, but the process is constrained by a low learning rate and gradient norm clipping to prevent catastrophic forgetting and ensure stability.

3. Key Contributions

Novel Architecture: Introduction of the GEGLU-Transformer for IMU-to-EMG estimation, which outperforms state-of-the-art recurrent baselines (LSTM and CNN-LSTM) in cross-subject generalization.
Robustness to Environmental Variability: Systematic evaluation across diverse locomotion modes (level ground, stairs, ramps, treadmill), demonstrating that the model's superiority is not limited to simple walking conditions.
Efficient Few-Shot Adaptation: Demonstration that the proposed architecture achieves rapid personalization with minimal data (as little as 0.5% of a subject's gait cycles), reaching performance saturation quickly compared to baselines.

4. Experimental Results

The model was evaluated using a Leave-One-Subject-Out (LOSO) cross-validation protocol on a dataset of 22 subjects.

Baseline Comparison (No Adaptation):
- Correlation ( $r$ ): GEGLU-Transformer achieved 0.706 (vs. 0.664 for CNN-LSTM and 0.647 for LSTM).
- Coefficient of Determination ( $R^2$ ): GEGLU-Transformer achieved 0.474 (vs. 0.420 and 0.397).
- Error Metrics: It showed lower Normalized Root Mean Square Error (nRMSE) and reduced peak timing/amplitude errors.
- Note: The GEGLU variant specifically improved $R^2$ by ~1.5% and correlation by ~0.33% over a non-gated Transformer, confirming the benefit of the gating mechanism.
Few-Shot Adaptation (0.5% Data):
- With only 0.5% adaptation data, the GEGLU-Transformer's correlation increased to 0.761 and $R^2$ to 0.559.
- This represents a significant jump in performance with negligible data, whereas baselines showed much smaller gains.
- Performance gains saturated rapidly; increasing data beyond 5% yielded marginal improvements, indicating high sample efficiency.
Muscle-Specific Performance:
- The model performed best on high-amplitude, regular muscles (e.g., Soleus, Gastrocnemius, Vastus).
- Lower-performing muscles (e.g., Gracilis, Biceps Femoris) showed higher variability, suggesting challenges related to signal quality and biomechanical complexity rather than just model capacity.

5. Significance and Impact

Scalability for Wearable Robotics: By reducing the reliance on direct EMG sensors (which are fragile and hard to calibrate) and enabling rapid personalization with minimal data, this approach makes EMG-driven control feasible for real-world, long-term use.
Adaptive Control: The framework supports the "adaptive control" paradigm in robotics, allowing systems to quickly adjust to new users without extensive calibration sessions.
Generalizability: The task-agnostic nature of the architecture suggests it can be extended to upper-limb movements or other assistive robotic applications beyond lower-limb gait.
Efficiency: The results prove that attention-based architectures, when combined with gated units and lightweight adaptation, offer a superior alternative to traditional recurrent models for continuous neuromuscular regression.

In conclusion, this work provides a practical, scalable solution for estimating muscle activation from inertial sensors, bridging the gap between laboratory-grade accuracy and real-world deployability in wearable robotics.