💻 computer science

Event-Driven On-Sensor Locomotion Mode Recognition Using a Shank-Mounted IMU with Embedded Machine Learning for Exoskeleton Control

This paper presents a low-latency, energy-efficient wearable system that utilizes an embedded machine learning core within a shank-mounted IMU to perform real-time, interrupt-driven recognition of locomotion modes (stance, walking, and stair ascent) directly at the sensor level, thereby minimizing host microcontroller power consumption and communication overhead for lower-limb exoskeleton control.

Original authors: Mohammadsaleh Razmi, Iman Shojaei

Published 2026-02-26

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Mohammadsaleh Razmi, Iman Shojaei

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 wearing a high-tech exoskeleton (a robotic suit) on your legs to help you walk or climb stairs. For this robot to help you effectively, it needs to know exactly what you are doing right now. Are you standing still? Walking on flat ground? Climbing up a flight of stairs?

If the robot guesses wrong, it might push your leg when you're trying to stop, or fail to help when you need it.

This paper describes a clever new way to teach the robot's "brain" to figure this out instantly, without draining the battery or slowing things down.

The Old Way: The Overworked Secretary

In the past, imagine the sensor on your leg (an IMU) was like a nervous secretary who never stops talking. Every single second, it would shout out raw data to the main computer: "My arm moved left! My arm moved right! The angle changed! The speed increased!"

The main computer (the microcontroller) would have to listen to this constant chatter, stop what it was doing, and try to figure out, "Okay, based on all this noise, is he walking or climbing?"

The Problem: This takes a lot of energy (like a phone battery dying fast) and creates a delay. By the time the computer figures it out, you might have already started climbing the stairs.

The New Way: The Smart Detective on the Scene

The authors of this paper came up with a smarter idea. Instead of sending all the raw data to the main computer, they put a tiny, super-smart detective inside the sensor itself.

Here is how it works, using a simple analogy:

The Sensor is the Detective: They took a specific sensor (the LSM6DSV16X) and installed a tiny, pre-trained "brain" (a Machine Learning model) directly inside it. Think of this as giving the sensor its own little library of rules.
The Training: They taught this sensor-detective by showing it thousands of examples of people standing, walking, and climbing stairs. The detective learned the "fingerprints" of each activity. For example, it learned that "climbing stairs" feels like a specific rhythm of shaking, while "walking" feels like a smooth, rolling wave.
The Decision: Now, when you move, the sensor-detective does all the thinking on its own. It doesn't shout out raw data. It just whispers a single word to the main computer: "I'm walking," or "I'm climbing."
The Sleepy Computer: Because the sensor does the hard work, the main computer can stay asleep (in a low-power state) most of the time. It only wakes up when the sensor taps it on the shoulder (an interrupt) to say, "Hey, I figured out what you're doing!"

Why This is a Big Deal

Speed (Low Latency): Because the thinking happens right where the movement happens, the robot reacts almost instantly. It's like a reflex; your knee jerks before your brain even thinks about it.
Battery Life: The main computer doesn't have to work hard, so the battery lasts much longer. It's like switching from a roaring engine to an electric motor that only kicks in when needed.
Simplicity: The "brain" inside the sensor is a very simple "Decision Tree." Imagine a flowchart:
- Is the leg shaking a lot? Yes -> Is it moving up and down quickly? Yes -> Climbing Stairs!
- Is it moving side-to-side smoothly? Yes -> Walking!
- Is it still? Yes -> Standing!
  This simple logic is so efficient it fits perfectly inside the tiny chip.

The Experiment

The researchers tested this by strapping the sensor to people's shins (lower legs). They made the people stand still, walk on a treadmill, and climb a stair machine.

The Result: The sensor-detective got it right 100% of the time during the test. It correctly switched from "Standing" to "Walking" to "Climbing" without any confusion.
The Proof: The main computer just sat there, sleeping, and woke up only to read the final answer. No heavy lifting required.

The Bottom Line

This paper shows that we don't need a super-computer to control a robotic suit. We can put the intelligence right inside the sensor. It's like upgrading a smoke detector: instead of sending a video feed of the kitchen to a central station to analyze if there's smoke, the detector itself smells the smoke and just rings the alarm.

This makes wearable robots faster, smarter, and much more energy-efficient, paving the way for better assistive devices for people who need them.

1. Problem Statement

Lower-limb exoskeletons require accurate, low-latency recognition of user activity (Human Activity Recognition or HAR) to enable smooth transitions between locomotion modes (e.g., standing, walking, climbing stairs) and to provide adaptive, context-aware torque assistance.

Current Limitations:

High Power Consumption: Conventional HAR systems continuously stream raw inertial data from the sensor to a host microcontroller for processing. This constant data transfer and computation drain battery life.
System Complexity: Offloading processing to the host increases communication overhead and requires complex software pipelines on the main controller.
Latency: Continuous streaming and external processing can introduce delays that hinder real-time control loops essential for safe and responsive exoskeleton operation.

2. Methodology

The authors propose a sensor-level inference architecture where activity recognition is performed entirely within the inertial measurement unit (IMU), eliminating the need for the host microcontroller to process raw data.

Hardware Setup:

Sensor: STMicroelectronics LSM6DSV16X IMU (integrated 3D accelerometer and 3D gyroscope) mounted on the lateral aspect of the participant's shank (lower leg).
Platform: ST SensorTile.box (STEVAL-MKSBOX1V1).
Sampling Rate: 7.68 kHz (raw data collection), with the embedded Machine Learning Core (MLC) operating at 240 Hz.

Software & Algorithm Development:

Toolchain: Development was conducted using ST MEMS Studio.
Data Collection: Adult participants performed three activities: Stance (quiet standing), Level Walking (treadmill), and Stair Ascent (step climber).
Feature Engineering:
- The Activity Feature Selection (AFS) tool was used to identify discriminative features from the inertial channels.
- 15 Features were selected, including variance, energy, peak-to-peak values, means, and extrema for both accelerometer and gyroscope axes (e.g., VARIANCE_GYR_X, PEAK_TO_PEAK_ACC_X).
Model Architecture:
- A lightweight Decision Tree classifier was trained.
- The final model is extremely compact (3 leaves, size 5), capable of classifying three classes: stance, walk, and stairsUp.
Deployment: The trained model was exported as an MLC configuration and deployed directly to the IMU's embedded Machine Learning Core.

Operational Workflow:

The IMU continuously processes raw data internally using the embedded MLC.
When a new classification is detected or motion is identified, the IMU asserts an interrupt (via INT1).
The host microcontroller remains in a low-power sleep state until the interrupt occurs.
Upon waking, the host reads only the activity label from the IMU's output registers.
The host forwards this label to the exoskeleton controller for torque adjustment.

3. Key Contributions

On-Sensor Inference Architecture: Demonstrated a fully functional HAR system where classification occurs entirely within the sensor hardware, removing the need for custom ML code on the host microcontroller.
Event-Driven Power Efficiency: By shifting from continuous data streaming to an interrupt-driven model, the system significantly reduces host CPU duty cycle and communication overhead, preserving battery energy.
Compact Model Design: Successfully deployed a highly interpretable, ultra-lightweight decision tree (3 leaves) that achieves high accuracy while fitting within the strict memory constraints of the IMU's MLC.
Robust Discrimination: The system effectively distinguishes between level walking and stair ascent, a critical challenge for lower-limb exoskeletons requiring different torque profiles.

4. Results

Training Performance: The model achieved 100% accuracy (43/43 instances) on the training dataset with a Cohen's $\kappa$ of 1.0. The confusion matrix showed perfect separation between walk, stairsUp, and stance.
Real-Time Evaluation:
- In continuous trials involving transitions from stance $\to$ walking $\to$ stair ascent, the system produced stable mode estimates.
- The host microcontroller successfully operated as a passive, interrupt-driven reader, logging mode changes without performing any signal processing or feature extraction.
- Latency: The system provided immediate mode updates via the MLC output registers, suitable for real-time control loops.
Signal Characteristics: The decision tree relied on physiologically plausible features, such as peak-to-peak acceleration (distinguishing dynamic motion from stance) and gyroscope mean values (distinguishing angular patterns of walking vs. stair climbing).

5. Significance and Future Work

Significance:
This work represents a paradigm shift from host-centric to sensor-autonomous wearable intelligence. By embedding machine learning directly into the sensing hardware, the system offers a scalable, low-power solution for "always-on" context awareness in assistive robotics. It simplifies system integration by providing a direct mode signal rather than a raw data stream, making it ideal for battery-constrained wearable devices.

Limitations & Future Directions:

Generalization: Current results are based on a specific training set; broader validation across different subjects, walking speeds, and sensor placement variations is required.
Metrics: Future work needs to quantify transition delays and event-level accuracy using synchronized ground truth (e.g., video).
Expansion: Plans include adding more locomotion modes (e.g., stair descent) and integrating the system into a closed-loop exoskeleton controller to evaluate the impact on user comfort, safety, and assistance timing.

In conclusion, this paper validates the feasibility of using embedded Machine Learning Cores in commercial IMUs to create efficient, low-latency control systems for next-generation lower-limb exoskeletons.