Characterizing Expectation Mismatch in a Brain-Controlled Upper-Body Rehabilitation Exoskeleton

Satyam Kumar1, Kanishka Mitra1, Deland H. Liu1, Hussein Alawieh1, Frigyes Samuel Racz2, Stefano Dalla Gasperina3, Ashish D. Deshpande3,5, José del R. Millán1,2,4
1 Chandra Family 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 4 Department of Biomedical Engineering, The University of Texas at Austin 5 Meta Reality Labs Research RA-L, 2025
Paper IEEE
Representative figure for the expectation mismatch paper

Error-aware rehabilitation control: A brain-controlled upper-body exoskeleton study that transfers an expert decoder to naive users and characterizes expectation-mismatch ErrPs during closed-loop control.

Abstract

Robot-assisted therapy can deliver intensive training, but its clinical impact remains limited when the user is only passively moved by the robot. This paper closes the loop with a motor-imagery BCI and studies what happens when the robot’s behavior does not match the user’s intent. The authors show that a decoder trained on an expert subject can be transferred to naive users for online control of an upper-body rehabilitation exoskeleton in a rest-versus-reaching task. They then characterize error-related potentials generated by expectation mismatch between brain commands and robot actions during closed-loop control, and show that those mismatches can be decoded in a subject-independent setting with a mean AUC of 0.77.

Experimental Setup and Closed-Loop Paradigm

Subjects used a motor-imagery BCI to control Harmony, an upper-body rehabilitation exoskeleton, in a rest-versus- reaching task. Successful reaching imagery drove the robot toward a goal, successful rest imagery kept it still, and failures to sustain the intended state triggered erroneous robot behaviors. Those mismatches between expectation and robot action created the ErrP events studied in the rest of the paper.

Placeholder for the experimental setup figure

Figure 1. Experimental setup. Subjects operated an MI-BCI to control Harmony, an upper-limb exoskeleton, in reaching for objects. Successful delivery of reaching MI drove Harmony toward the goal, while failures to sustain the intended state produced erroneous robot behaviors.

Grand Average ErrPs

The paper first establishes the expectation-mismatch signal directly in the EEG. Grand average erroneous and correct trials are compared at FCz, with the mismatch event aligned to time zero and spatial insets showing the topography near the positive and negative peaks of the ErrP response.

Placeholder for the grand average ErrPs figure

Figure 2. Grand average ErrPs. Time-locked grand averages of erroneous and correct classes across subjects (n = 8) at FCz, with shaded regions showing the standard error of the mean and topography insets centered on the main ErrP peaks.

Results

The results show two linked outcomes. First, naive users can operate the exoskeleton online using an expert-trained decoder, making closed-loop error events available without subject- specific calibration. Second, those expectation-mismatch events produce clear neurophysiological signatures that support subject-independent ErrP decoding, a practical requirement for rehabilitation scenarios where collecting personalized ErrP datasets is difficult.

Placeholder for the neural markers of expectation mismatch figure

Figure 3. Neural markers of expectation mismatch. Each dot corresponds to the subject-level average marker for correct and erroneous classes, including P1, N2, and theta-power ratio comparisons.

Placeholder for the time-frequency ErrP decomposition figure

Figure 4. Time-frequency decomposition of ErrPs. Difference in time-frequency representations between erroneous and correct trials across subjects, highlighting theta-band power changes relative to event onset at FCz.

Toward Subject-Independent Error Detection

The key practical contribution is that expectation mismatch can be decoded without collecting subject-specific ErrP training data for every user. That combination of calibration-free MI control and subject-independent error detection points toward more reliable brain-controlled rehabilitation robots, where erroneous actions can eventually be detected and corrected online.

BibTeX

@article{kumar2025characterizing,
  title={Characterizing Expectation Mismatch in a Brain-Controlled Upper-Body Rehabilitation Exoskeleton},
  author={Kumar, Satyam and Mitra, Kanishka and Liu, Deland H and Alawieh, Hussein and Racz, Frigyes Samuel and Dalla Gasperina, Stefano and Deshpande, Ashish D and Mill{\'a}n, Jos{\'e} del R},
  journal={IEEE Robotics and Automation Letters},
  volume={11},
  number={1},
  pages={762--769},
  year={2025},
  publisher={IEEE}
}