Characterizing the onset and offset of motor imagery during passive arm movements induced by an upper-body exoskeleton

Kanishka Mitra1, Frigyes Samuel Racz2, Satyam Kumar1, Ashish D. Deshpande3, José del R. Millán1,2
1 Chandra Department of Electrical and Computer Engineering, The University of Texas at Austin 2 Department of Neurology, The University of Texas at Austin 3 Walker Department of Mechanical Engineering, The University of Texas at Austin IROS 2023
Paper IEEE
Representative figure for the passive arm motor imagery paper

Onset and offset under passive motion: An EEG decoding framework that studies when kinesthetic motor imagery begins and ends while the arm is being moved by a rehabilitation exoskeleton.

Abstract

This paper studies whether motor intentions can still be detected with noninvasive EEG when a rehabilitation exoskeleton introduces passive arm motion, instrumental noise, and additional sensorimotor activity. Instead of treating motor imagery as a single sustained state, the work separates movement initiation and termination into distinct decoding problems. Ten participants performed right-arm kinesthetic motor imagery while attached to the Harmony upper-body exoskeleton, with LEDs cueing the beginning and end of a goal-directed reach. Offline decoders distinguished both the transition from rest to motor imagery and the transition from sustained imagery to stopping, reaching group-average accuracies of 60.7% for onset and 66.6% for offset. Pseudo-online evaluation further suggested that reliable exoskeleton-compatible control is feasible, showing that participants can still produce informative sensorimotor rhythms despite passive movement and robot-induced noise.

Experimental Setup and Protocol

Ten healthy participants performed kinesthetic motor imagery of the right arm while wearing the Harmony upper-body exoskeleton. The task cued the initiation and termination of a goal-oriented reaching movement with LEDs, allowing the study to probe both the beginning and ending of imagery in a setting where passive arm motion and exoskeleton-induced activity are present in the EEG recordings.

Representative setup figure for the passive arm motor imagery paper

Figure 1. Experimental setup and task paradigm pipeline, including the Harmony exoskeleton setup and the trial timeline.

Spectral Dynamics During the Task

The paper first characterizes how spectral power evolves over the trial, showing event-related desynchronization and robot-related changes in the C3 region. This figure sets up the neural context for the onset and offset decoding tasks by showing when task-relevant activity emerges relative to rest, motor imagery, and passive exoskeleton motion.

Grand average spectrogram placeholder for the C3 region

Figure 2. Grand average spectrogram from the C3 region showing changes in spectral power over the course of task trials.

Results

Offline evaluation showed that onset and offset can be decoded separately even when the exoskeleton introduces passive motion and noise. Group-average accuracy reached 60.7% for onset and 66.6% for offset, suggesting that the start and stop of motor imagery remain distinguishable while attached to the robot. The results section also analyzes how classification evolves over time within the one-second decoding windows and which frequency features are most informative for onset versus offset.

Grand average classification accuracies over time placeholder

Figure 3. Grand average classification accuracies over time during the 1-second epochs for the onset and offset decoding tasks.

Top discriminative feature frequencies placeholder

Figure 4. Top 10 most discriminative feature frequencies. The color indicates how frequently each feature was selected for MI-onset and MI-offset distinctions.

Control Trials and Pseudo-Online Evaluation

The later figures validate that the observed patterns are not driven only by passive movement or robot noise, and they also show that a Riemannian pseudo-online decoder can track onset and offset transitions continuously over time. Together, these figures are the bridge from offline signal characterization toward future online exoskeleton control.

Control trial spectrogram comparison placeholder

Figure 5. Average spectrograms for the two re-recorded subjects with control trials, comparing test and control conditions.

Riemannian classifier time-resolved performance placeholder

Figure 6. Time-resolved performance of the Riemannian classifier across the relevant onset and offset trial segments.

BibTeX

@inproceedings{mitra2023characterizing,
  title={Characterizing the onset and offset of motor imagery during passive arm movements induced by an upper-body exoskeleton},
  author={Mitra, Kanishka and Racz, Frigyes Samuel and Kumar, Satyam and Deshpande, Ashish D and Mill{\'a}n, Jos{\'e} Del R},
  booktitle={2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  pages={3789--3794},
  year={2023},
  organization={IEEE}
}