Performing Bimanual Tasks with a BCI: Combining a Brain-Controlled Hand Exoskeleton with the Functional Limb

Satyam Kumar1, Kanishka Mitra1, Ruofan Liu1, Hussein Alawieh1, Akhil Surapaneni2, Ashish D. Deshpande3,4, José del R. Millán1,2,5
1 Chandra Family Department of Electrical and Computer Engineering, The University of Texas at Austin 2 Department of Biomedical Engineering, The University of Texas at Austin 3 Walker Department of Mechanical Engineering, The University of Texas at Austin 4 Meta Reality Labs Research 5 Department of Neurology, The University of Texas at Austin International Conference on Neural Engineering (NER), 2025
Paper arXiv
Preview of the bimanual BCI paper

Bimanual assistive control with a transferred decoder: A multi-day study showing that a unimanual hand-flexion MI-BCI decoder can transfer to realistic bimanual tasks with a brain-controlled hand exoskeleton.

Abstract

Brain-controlled robotic systems have shown promise as assistive tools, but most work still focuses on single-arm exoskeleton use rather than the bimanual coordination needed for everyday tasks. This paper studies whether a decoder trained on a simpler unimanual hand-flexion versus rest task can be transferred to more realistic bimanual behavior, where a user controls a robotic hand exoskeleton with MI while also using the functional limb. Across five sessions with seven participants, the study shows that transfer is feasible, multi-day closed-loop training improves performance, and the resulting feature space becomes more discriminative over time.

Experimental Setup and Bimanual Task Design

Subjects controlled a right-hand assistive exoskeleton with a motor-imagery BCI while simultaneously performing a left-hand functional task. The study began with offline and online unimanual training, then transferred the same flexion-versus-rest decoder to four synchronous bimanual conditions that combined functional-hand opening or closing with BCI-driven exoskeleton opening or closing.

Bimanual experimental setup and task conditions

Figure 1. Bimanual experimental setup. Subjects are cued to simultaneously control the assistive robotic exoskeleton using MI-BCI while using the functional left hand to apply force or rest using a cylindrical grip. The subpanels show the four bimanual task conditions used during online evaluation.

Online BCI Performance

The paper tracks sample-level online MI classification performance across unimanual and bimanual sessions. Although performance drops initially when subjects transition to the more demanding simultaneous-control setting, it improves substantially with repeated training across days.

Bimanual BCI control across unimanual and bimanual sessions

Figure 2. Bimanual BCI control. Session-level kappa value of MI classification performance for each unimanual and bimanual online session. Different markers represent individual subjects (n = 7).

Feature Separability and Decoder Adaptation

Two complementary analyses explain how control improves. First, the Riemannian feature space becomes more distinctive over the course of training, suggesting that subjects learn to generate more separable neural patterns. Second, a pseudo-online comparison shows why the adaptive decoder matters: the incrementally recentered online decoder consistently outperforms a fixed decoder that does not correct for non-stationarities.

Feature separability analysis across unimanual and bimanual sessions

Figure 3. Feature separability analysis. Session-level Riemannian distinctiveness illustrating the separability of feature spaces in (a) unimanual online sessions and (b) bimanual online sessions. Different markers represent individual subjects (n = 7).

Pseudo-online comparison of fixed and adaptive decoders

Figure 4. Pseudo-online decoder comparison. Bar plot comparing the fixed decoder and the adaptive online decoder across unimanual and bimanual sessions, with mean kappa values, standard deviations, and paired-test p-values.

Event-Related Desynchronization Patterns

The paper also visualizes the sensorimotor rhythms that accompany successful control. Unimanual trials show the expected contrast between rest-related synchronization and flexion-imagery desynchronization, while the bimanual plots reveal how simultaneous use of the functional hand changes the resulting topographies.

Event-related desynchronization during the unimanual online session

Figure 5. Event-related desynchronization during the unimanual online session. Grand-average topographies of mu-band ERD across subjects and unimanual sessions for (a) rest imagery and (b) right-hand flexion imagery.

Event-related desynchronization during the bimanual online session

Figure 6. Event-related desynchronization during the bimanual online sessions across the four combinations of functional-left-hand and BCI-right-hand conditions: open/rest, closed/rest, open/closed, and closed/closed.

BibTeX

@inproceedings{kumar2025bimanual,
  title={Performing Bimanual Tasks with a BCI: Combining a Brain-Controlled Hand Exoskeleton with the Functional Limb},
  author={Kumar, Satyam and Mitra, Kanishka and Liu, Ruofan and Alawieh, Hussein and Surapaneni, Akhil and Deshpande, Ashish D and Mill{\'a}n, Jos{\'e} del R},
  booktitle={International Conference on Neural Engineering (NER)},
  year={2025},
  organization={IEEE}
}