Work

EEG Test-Time Adaptation in Offline Robot Simulation: State-Contingent Replay Without a User in the Loop

Year: 2026
Type: article
Abstract: EEG-driven control couples decoder predictions, robot state and subsequent EEG requests. We introduce an offline, simulation-only framework for state-contingent replay of prerecorded EEG with no user in the loop. Shared within-class schedules support paired policy comparisons, while component matching distinguishes semantic adaptation from geometry assistance. Safe-DPPC combines source-anchored prototypes, a running class prior and output selection around a frozen source network. Findings from the nine-participant BNCI2014-001 development cohort are descriptive. With shared source checkpoints, buffered batch normalization (BN32) and buffered Tent (Tent32) achieved balanced accuracy of 49.20% and 48.15%, respectively, versus 42.60% for Source and 24.98% for single-trial Tent; both buffered methods had lower mean navigation success than Source in the two route-reset conditions. Under three-trial route-reset requests, Tent32 increased strict confirmation by 18.11 percentage points relative to single-trial Tent, with more requests. Relative to Source, Safe-DPPC showed little change in common-input balanced accuracy, higher calibration error and a −2.26-percentage-point change in primary natural-condition navigation success. GeoShield reduced mean collision exposure by 11.07 percentage points in that condition while preserving 98.02% of command classes. None of the four primary contrasts reached Holm-adjusted significance. The framework relates update timing and retained evidence to decoding, arrival, confirmation and request burden along the resulting trajectories.
Source: Electronics
Language: English
Open Access status: gold
Estimated APC paid: $2,896
OpenAlex ID: W7233742902
FWCI:
Cites: 52
Cited by:
Related to: 10