Segmentation and Classification of EMG Time-Series During Reach-to-Grasp Motion

Mo Han, Mehrshad Zandigohar, Mariusz P. Furmanek, Mathew Yarossi, Gunar Schirner, Deniz ErdogmusarXiv 2021 · arXiv

Introduction

In the field of human-robot interaction (HRI), particularly for assistive robotics and smart prosthetics, the ability to predict a user's intent in real-time is paramount. Current systems often rely on "steady-state" classification, which recognizes a gesture only after it has been fully formed. However, natural human movement is fluid; for instance, the hand begins "pre-shaping" during the reach phase long before contact with an object. This paper presents a framework for predicting grasp intent during this dynamic "reach-to-grasp" phase using electromyography (EMG) signals.

The Challenge of Dynamic Intent Prediction

Capturing intent during fluid movement presents several hurdles. Traditional labeling requires manual timestamping or expensive kinematic sensors like data gloves to distinguish between movement phases. Furthermore, EMG signals exhibit high variability during active movement compared to static poses. Finally, for a robotic system to be effective, it must identify intent early enough to prepare its own configuration to match the user's upcoming grasp.

Methodology

The researchers developed an automated pipeline to address these challenges without relying on supervised kinematic labeling.

Unsupervised Segmentation

The framework utilizes Greedy Gaussian Segmentation (GGS) to automatically partition the EMG data stream. By treating the signal as a series of multivariate Gaussian distributions, GGS identifies "breakpoints" where statistical properties shift, allowing the system to segment trials into four distinct phases: reach, grasp, return, and rest.

Feature Extraction and Classification

To ensure low computational overhead for real-time applications, three time-domain features were extracted: Root Mean Square (RMS), Mean Absolute Value (MAV), and Variance (VAR). These features were calculated over 320ms sliding windows. An Extra-Trees (Extremely Randomized Trees) classifier, consisting of 50 trees, was then employed to identify specific gesture labels.

Experimental Setup

The study involved five healthy subjects performing a vocabulary of 14 gestures, including 13 distinct grasps from the Feix taxonomy and a rest state. A total of 52 different objects were used to provide variety in grasp requirements.

Electrode placement for EMG signal acquisition across anterior, lateral, and posterior views of the arm.

As shown in Figure 1, EMG sensors were placed across the forearm and hand to capture the muscular signatures of the reaching and grasping phases. Each subject performed hundreds of trials, which were validated using a 3-fold cross-validation protocol.

Results and Key Insights

The study's most significant finding is the existence of a reliable "pre-shaping" window for intent inference.

  • Early Detection: By training on the reaching phase, the system can identify the intended grasp approximately 740ms before the actual grasp occurs.
  • Informative Reaching Phase: The results demonstrate that transient EMG signals during the reach phase contain sufficient unique signatures to differentiate complex grasps, challenging the notion that peak muscle contraction is required for identification.
  • Classification Accuracy: The system achieved an average accuracy of 85% during the critical transition from reaching to grasping.
  • Confidence and Robustness: Models trained on both reach and grasp phases yielded the highest confidence, with a significant probability distance between the correct gesture and competitors.

Conclusion

The proposed unsupervised framework successfully identifies grasp intent well before physical contact, facilitating smoother human-robot collaboration. While the study was limited to healthy subjects, it lays the groundwork for future applications in prosthetics. Future work will focus on validating these models with amputee populations and refining the segmentation of the reach phase to further improve lead times and classification granularity.

Summary updated

BibTeX
@misc{han2021segmentation,
  author = {Mo Han and Mehrshad Zandigohar and Mariusz P. Furmanek and Mathew Yarossi and Gunar Schirner and Deniz Erdogmus},
  title = {{Segmentation and Classification of EMG Time-Series During Reach-to-Grasp Motion}},
  year = {2021},
  eprint = {2104.09627},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2104.09627},
}
Figure