Google Drive Link
In the last post I argued that the research had to stop being a broad map and become one concrete thing I could build and test. This is that thing.
I built a neuroadaptive supervision console. On screen, the operator keeps a small fleet of semi-autonomous drones alive: each one occasionally needs attention and timed emergencies appear that must be resolved before a countdown runs out. The drones are not the point for me, they are a stand-in for any complex, information-dense interface, the kind of control-room or monitoring setting where alerts compete for attention and overload is a constant risk. That is where badly-timed interruptions do the most damage, so that is the situation I wanted to recreate.
Underneath the task, the system builds a live estimate of how loaded the operator is, by combining three real signals: EEG engagement from a Muse headband, a webcam-based attention proxy (for blink rate and head movement – unfortunately I could not extract data from Tobii) and the demand of the task itself. I then compared two versions of the same interface. In the adaptive version, when the estimate crosses into “overload,” the system quietly defers low-urgency alerts until there is room for them and steps in for the drones most at risk. In the control version it does none of that. Every participant played both.
The clearest result, shown in the video, is reaction time. On the working build, people answered emergencies far faster in the adaptive version – around four seconds, against seventeen in the control. The mechanism is simple: when alerts are not deferred, they pile up during busy moments and get answered at the last second. Deferring them spreads the load out.
What the video does not point out on is that this was not a clean win and I think that matters more than the headline. Faster answers came with lower accuracy, and people did not report feeling less loaded. With only four participants, and a build that changed between the early and late sessions, this is a proof-of-concept, not proof. The fuller per-person picture, including a resumption-lag result that went the “wrong” way, sits in my written results rather than the video.
For me the takeaway is not the number but the shape of the problem. A real adaptive loop is buildable with modest, honest signals, and once it works the interesting questions are not about better sensors. They are about timing, trust and giving the person a way to overrule the system when its guess is wrong. That is the thread I want to keep pulling: from interfaces that capture attention toward interfaces that look after it.
