On the Feasibility of EEG-based Motor Intention Detection for Real-Time Robot Assistive Control

TL;DR

The paper tackles real-time EEG-based motor intention detection to drive assistive robots. It introduces a Riemannian-geometry–driven framework that projects time-windowed covariance matrices to the tangent space, producing 465-dimensional features for an SVM classifier, and validates offline real-time performance up to 86.88% at 160 Hz with a 2 s window. In robot-in-the-loop tests with two subjects, it achieves about 70% accuracy, aided by score thresholding and a buffering strategy to stabilize control, demonstrated on a KUKA iiwa 7. The results show that high-accuracy, low-data EEG decoding is feasible for real-time assistive control, with potential extensions to EMG signals and more complex motor tasks.

Abstract

This paper explores the feasibility of employing EEG-based intention detection for real-time robot assistive control. We focus on predicting and distinguishing motor intentions of left/right arm movements by presenting: i) an offline data collection and training pipeline, used to train a classifier for left/right motion intention prediction, and ii) an online real-time prediction pipeline leveraging the trained classifier and integrated with an assistive robot. Central to our approach is a rich feature representation composed of the tangent space projection of time-windowed sample covariance matrices from EEG filtered signals and derivatives; allowing for a simple SVM classifier to achieve unprecedented accuracy and real-time performance. In pre-recorded real-time settings (160 Hz), a peak accuracy of 86.88% is achieved, surpassing prior works. In robot-in-the-loop settings, our system successfully detects intended motion solely from EEG data with 70% accuracy, triggering a robot to execute an assistive task. We provide a comprehensive evaluation of the proposed classifier.

Figures (5)

• Figure 1: Real-time EEG Motor Intention Prediction Framework. In the offline pipeline, we first gather 4 minutes of raw EEG data as well as arm kinematics given by the motion capture system used to define the corresponding classification labels (left vs. right arm motion). In both offline and online pipelines we use the same feature extraction approach described in Section \ref{['sec:methods']} to process the EEG signals and construct a feature vector as the tangent space projection of the sample covariance matrices. An SVM classifier is trained per subject on 50% of the collected data. The trained SVM classifier is then used online to predict left vs. right arm motion intention solely from EEG data, due to change of setting with the robot-in-the-loop a score thresholding phase is added to the pipeline to filter out noisy prediction and trigger the robot to provide assistance in reaching for an object with the left vs. right arm with high confidence.
• Figure 2: Illustration of a Riemannian manifold and the tangent space of a point. (left) Tangent space projection from $P^*$ to $S^*$ using $P$'s tangent space. (right) A step of the iterative EM process used to determine the mean of points on the Riemannian manifold.
• Figure 3: Confusion matrix of robot-in-the-loop experiments
• Figure 4: Mean probability confidence for all left and right trials in black line with respect to the onset of motion and the readiness potential (RP) in red line in $\mu V$ units reproduced from Wen2018TheRP.
• Figure I: Mean accuracy of SVM classification on testing datasets across various frequency ranges (columns) and time windows (rows).