Abstract:
Motor imagery is a learned skill. An estimated 15–30 % of healthy adults are BCIilliterate
[1], and the cue-based paradigm that produces that number is known to be
tedious, monotonous, and fatigue-inducing long before a training session is over. This
thesis treats that as an engineering problem: if the training environment is what determines
whether a novice can learn motor imagery, then the training environment is where the
effort should go.
I built Brain Race, a real-time three-lane racing game driven by mu-band ERD
from C3 and C4 on a consumer-grade Bitbrain Versatile 32 headset. The game features a
40-second per-subject calibration that tunes the lane-switch threshold to the user's own
ERD amplitude, three session modes (classic, precision, endurance) with adaptive
difficulty, and a sub-100 ms electrode-to-pixel latency budget (stage-by-stage estimate,
chapter 3). Surrounding the game sit four supporting tools: a clean cue-based recorder for
classifier-training data, a live-ERD feedback interface for teaching a new user what motor
imagery feels like, a practice mode for setup and rehearsal, and a feedback recorder that
captures cue-aligned trials while the user sees a live ERD bar. There is also a per-trial 0–
100 quality scorer written live into every saved trial and shown to the player immediately
after each trial ends.
A classification pipeline runs under the game to turn the EEG signal into the live
feedback the player sees. I ran the pipeline on a public motor-imagery dataset as a leakfree
reference check and exercised it on two subjects' own recordings: the cross-paradigm
clean recorder and Brain Race. Across 30 sessions for participant_01 and 31 for
participant_22, 44 of 61 sessions cleared the chapter-5 per-trial quality threshold,
with 27 of those 44 produced during Brain Race play (out of 35 Brain Race sessions
total, versus 17 of 26 clean-recorder sessions). The same pattern held on the volunteer
who plateaued at classifier level under the cue-based paradigm. This is a within-subject
pattern on two subjects, not evidence that the game caused the cleaner recordings;
chapter 7 specifies what a controlled study would need to settle that. Full evaluation in
chapter 6.
Limitations are stated plainly: two subjects, unbalanced exposure, no randomised
control. The contribution is the game and the instrument around it, ready for the multisubject
controlled study that would be the proper test of gameplay- driven motor-imagery
training.