Out-of-the-lab robustness

The capacity of a BCI to keep working when the controlled conditions of the laboratory are removed: uncontrolled electromagnetic environments, imperfect and shifting electrode contact, the user in motion or multitasking, ambient distraction, and calibration data that no longer match the moment of use. It is distinct from accuracy on a benchmark; a system can be state-of-the-art offline and fragile in the world.

Framed in machine-learning terms, the core difficulty is distribution shift: the statistics of the deployed signal differ from those of the training set, and they differ in ways that were not sampled. Robustness is therefore engineered rather than assumed — through artefact-tolerant hardware and montages, features chosen for stability rather than peak separability, on-line adaptation and self-recalibration, uncertainty-aware decoders that can abstain when contact degrades, and graceful failure modes that ask for a re-fit instead of emitting confident nonsense.