Healthy adults inside an fMRI scanner learned to control a video game avatar with their brains in under an hour, when the decoder followed each person’s individual brain-activity patterns. When the decoder asked for patterns the brain does not naturally produce, the same participants could not learn the task at all in the same window.
The Yale study was published on 9 June 2026 in Nature Neuroscience. Erica Busch, who recently completed her PhD at Yale, is first author. The corresponding authors are Smita Krishnaswamy, associate professor of genetics at Yale School of Medicine and of computer science at Yale Engineering, and Nick Turk-Browne, director of the Wu Tsai Institute and the Susan Nolen-Hoeksema Professor of Psychology at Yale FAS. The work was funded by the US National Science Foundation, the Canadian Institute for Advanced Research, the Sloan Foundation, and the National Institutes of Health.
How the system works
Participants played a video game inside an fMRI scanner during a first session, steering an avatar through a virtual arena using a joystick while the researchers recorded brain activity. The team focused on a network of brain regions known to support spatial navigation. Using an algorithm Krishnaswamy’s group developed in earlier work called T-PHATE, the researchers extracted the natural geometry of each participant’s brain activity, the individual “neural manifold” along which that user’s activity tends to travel.
The team then built a closed-loop system that read a new fMRI scan every two seconds and translated it into avatar movement in real time. Three decoder mappings were tested, one per subsequent training session per participant. The “intuitive mapping” followed the most well-travelled neural route on that user’s manifold. The “within-manifold perturbation” used less dominant but still natural patterns. A third mapping, the “outside-manifold perturbation”, required brain patterns the user’s brain does not naturally produce.
What the participants learned
Participants learned avatar control in less than an hour, sometimes considerably faster, when the decoder followed their natural manifold. The off-manifold mapping produced no learning at all in the same timeframe. Brain reorganisation was observed during successful training, with activity shifting to better align with what the decoder required. The reorganisation spread to brain regions outside the originally targeted network, and in some conditions it predicted individual performance.
The baseline the study claims to beat
The Yale paper frames the historical fMRI-BCI field as the baseline this approach improves on. fMRI-based BCIs built on real-time neurofeedback have typically required up to ten long training sessions per participant, the paper reports, and about a third of users never gain control regardless of how many hours they practise. The framing offered by the authors is that those numbers reflect a decoder-design problem upstream of any user limitation. A decoder that asks the brain to learn patterns poorly matched to its natural geometry produces slow, unreliable learning. A decoder built around the natural geometry does not.
Why the result matters beyond Yale
Training cost is one of the largest barriers to commercial scale in BCI deployment. Any approach that compresses it by an order of magnitude changes the unit economics of every downstream use case, including clinical motor and communication BCIs, consumer-facing applications, and mental health interventions in which neurofeedback is itself the therapy. The Yale work also points at a substrate-agnostic principle. Decoders that match the user’s existing neural geometry are likely to outperform decoders that demand the user adapt to the decoder, regardless of whether the interface is fMRI, EEG, cortical implant, or peripheral neural interface. The principle is consistent with the direction the invasive BCI field has been moving, with multiple clinical-stage companies investing in adaptive decoder approaches that recalibrate to user activity rather than holding the decoder fixed.
Scope and limits
The study used healthy young adults and non-invasive fMRI, which sits at the methodology-advance end of the BCI literature rather than the clinical-milestone end. Whether the manifold-following approach holds in clinical populations with neural reorganisation following injury or disease is a separate question for the next study. The closed-loop fMRI platform itself is not a candidate for commercial deployment outside a hospital scanner. The transferable contribution is the decoder-design principle, not the hardware.
Busch and Turk-Browne both note in the Yale announcement that the implications extend beyond BCI usability. If learning depends on how well a task fits a person’s existing neural architecture, that has direct bearing on training, education, and clinical interventions for depression and anxiety, where the brain “gets stuck in unhelpful grooves” and treatment may work better when it follows existing routes rather than attempting full overhauls.
Sources