Real-Time Decoding of Movement Onset and Offset for Brain-Controlled Rehabilitation Exoskeleton

Kanishka Mitra1,2, Satyam Kumar2, Frigyes Samuel Racz2, Deland Liu2, Ashish D. Deshpande2,3, José del R. Millán2
1 Massachusetts Institute of Technology 2 The University of Texas at Austin 3 Meta Reality Labs Research ICRA 2026
Paper Code arXiv
Experimental setup, task timeline, recentering method, and decoder pipeline

Dual-state EEG control: A noninvasive brain-computer interface that starts and stops rehabilitation exoskeleton assistance in real time, while using fixation-based recentering to improve robustness across sessions.

Abstract

Robot-assisted therapy can deliver high-dose, task-specific training after neurologic injury, but most systems act primarily at the limb level, engaging the impaired neural circuits only indirectly. We address this gap by implementing online, dual-state motor imagery control of an upper-limb exoskeleton, enabling goal-directed reaches to be both initiated and terminated directly from noninvasive EEG. Eight participants used EEG to initiate assistance and then volitionally halt the robot mid-trajectory. Across two online sessions, group-mean hit rates were 61.5% for onset and 64.5% for offset, demonstrating reliable start-stop command delivery despite instrumental noise and passive arm motion. Methodologically, we reveal a systematic, class-driven bias induced by common task-based recentering using an asymmetric margin diagnostic, and we introduce a class-agnostic fixation-based recentering method that tracks drift without sampling command classes while preserving class geometry. This substantially improves threshold-free separability and reduces bias within and across days, helping bridge offline decoding and practical intention-driven control of a rehabilitation exoskeleton.

Experimental Setup and Task Timeline

Participants wore a 60-channel EEG cap while strapped into the Harmony upper-limb rehabilitation exoskeleton. Each trial cued a Start motor-imagery command to launch assistance toward one of three targets, followed by a Stop motor-imagery command to intentionally halt the reach mid-trajectory. Sensory-threshold neuromuscular electrical stimulation reinforced the current command state through agonist-antagonist pairing.

Experimental setup, task timeline, fixation-based recentering, and subject-specific decoding pipeline

Figure 1. Harmony setup, LED-cued task timeline, fixation-based recentering, and the subject-specific Riemannian decoding pipeline.

Neural Signatures During the Task

The offline spectrogram analysis shows clear mu-band desynchronization during Start motor imagery and a beta-band rebound after Stop motor imagery. These canonical signatures remain visible even with passive arm motion and exoskeleton-driven sensory feedback, supporting the use of dedicated onset and offset decoders rather than a single thresholded stream.

Grand-average C3 spectrogram with task events marked over time

Figure 2. Grand-average C3 spectrogram with task events marked over time.

Results

Online performance shows reliable dual-state control. Group-mean onset hit rate improved from 58% in Session 2 to 65% in Session 3, while offset hit rates were 66% and 63% on attempted trials. Onset decisions clustered near one second after the cue, whereas offset decisions arrived around 3.3 to 3.4 seconds after movement onset, consistent with participants intentionally timing the stop command close to the target.

Online hit, miss, and timeout rates for onset and offset across sessions

Figure 3. Hit, miss, and timeout rates for onset and offset across the two online sessions.

Onset and offset latency distributions for hit trials

Figure 4. Onset and offset latency distributions for hit trials only.

Fixation-Based Recentering

The paper shows that task-based recentering introduces a class-driven bias by shifting positive samples toward the pooled training center. Replacing that reference with a class-agnostic fixation period improves threshold-free AUC from 0.554 to 0.866 for onset and from 0.619 to 0.832 for offset, while reducing asymmetric margin shifts across runs and across days.

Margin-shift analysis showing bias introduced by task-based recentering

Figure 5. Margin-shift analysis for task-based recentering.

Run-wise AUC comparison for task-based versus fixation-based recentering

Figure 6. Run-wise AUC comparison for task-based versus fixation-based recentering.

BibTeX

@inproceedings{mitra2026realtime,
  title={Real-Time Decoding of Movement Onset and Offset for Brain-Controlled Rehabilitation Exoskeleton},
  author={Mitra, Kanishka and Kumar, Satyam and Racz, Frigyes Samuel and Liu, Deland and Deshpande, Ashish D. and Millan, Jose del R.},
  booktitle={IEEE International Conference on Robotics and Automation (ICRA)},
  year={2026},
  note={Add pages, DOI, and final publication URL}
}