Abstract
This study investigates the application of supervised learning algorithms to translate human electromyography (EMG) signals into discrete robotic arm positions. By comparing six different machine learning models, the research identifies k-Nearest Neighbors (kNN) and Random Forests as highly effective, achieving accuracies exceeding 97%. The findings suggest that standard machine learning techniques can provide the low-latency, intuitive control necessary for advanced telerobotic systems.
Introduction
Traditional Human-Computer Interaction (HCI) for teleoperation typically involves mechanical interfaces like joysticks or levers. While effective, these methods lack intuitiveness and often require extensive operator training. This research explores a more natural interface using EMG signals—the electrical manifestations of muscle activity. The primary challenge is the "mapping" problem: converting noisy, complex EMG data into precise mechanical commands in real-time. This study focuses on classifying static arm positions as a foundational step toward continuous motion control.
System Architecture and Methodology
The researchers developed an integrated pipeline to capture and process muscle signals for robotic control.
The hardware setup utilized surface electrodes placed on four muscle groups: the bicep, tricep, anterior deltoid, and posterior deltoid. Data was acquired through a NeuroFieldz prototype board and transmitted to a PC. Preprocessing involved a band-pass filter (50Hz – 400Hz) to eliminate DC offsets and motion artifacts, followed by rectification and the calculation of the Mean Absolute Value (MAV) for each muscle channel to serve as features.
Six supervised learning algorithms were evaluated:
- Naïve Bayes (NB)
- Support Vector Machines (SVM) with a radial basis function kernel
- Logistic Regression (LR)
- k-Nearest Neighbors (kNN) using L2 distance
- J48 Decision Trees
- Random Forests (RF)
The resulting commands were executed by a PhantomX Reactor Research Robot Arm via a microcontroller.
Data Collection
A custom dataset was generated from a single subject, comprising 369 instances across three distinct elbow angles: 90°, 150°, and a range of 100-110°. Ground truth labels were derived from the positional feedback of Dynamixel AX-12A robotic actuators, ensuring synchronization between human muscle state and robotic pose during the training phase.
Results and Discussion
The models were evaluated using 10-fold cross-validation. All algorithms demonstrated strong performance, with accuracies ranging from 90.0% to 97.8%.
As shown in the voltage distribution above, the MAV features for different positions form relatively distinct clusters, explaining why simple models like kNN (k=4) achieved the highest accuracy (97.8%).
| Algorithm | Accuracy (%) | Training Time (ms) | Testing Time (ms) |
|---|---|---|---|
| kNN (k=4) | 97.8% | 0 | 0.08 |
| Random Forest | 97.3% | 120 | 0.02 |
| SVM | 95.7% | 20 | 0.15 |
| J48 Tree | 95.1% | 20 | 0.02 |
| Naïve Bayes | 90.0% | 0 | 0.02 |
| Logistic Reg. | 90.1% | 30 | 0.03 |
Crucially, all models met the real-time requirement for telerobotics (latency < 50ms), with inference (testing) times under 1ms. While Random Forests were highly accurate, kNN offered a superior balance of performance and zero-cost training for this specific dataset size.
Conclusion and Future Work
This study confirms that EMG-based control is feasible using standard supervised learning. However, several challenges remain. The current model is subject-specific, and EMG signals vary significantly between individuals due to muscle mass and skin impedance. Future work will address generalization across multiple subjects and the transition from static position classification to continuous, fluid motion control, which involves managing more complex motion artifacts.
Summary updated