You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Assign a pilot to a performance group live, during a flight #223
Is your feature request related to a problem? Please describe. #216 clusters pilots in batch, after their flights are finished and logged. The client eventually wants to see which performance group a pilot falls into while the pilot is still flying. Nothing in inference/ reads the stream live or assigns a pilot to an existing performance group. The batch model is not saved either, so the only option today is to refit the whole cohort.
Describe the solution you'd like
A live mode that taps the same TCP stream inference/Data/data_logger.py reads. It computes deviation features for the flight in progress at a fixed refresh rate and assigns the pilot to the nearest performance group of a saved batch fit, without refitting.
Recompute the flight's deviation features (std and |bias| per flight dynamics factor) from all samples so far. Reuse inference/features.py, not a second implementation.
Scale them with the saved scaler and assign the nearest centre.
Withhold an assignment until the flight has enough samples to make it meaningful.
Open questions (needs-info)
Refresh rate: how often to reassign (every N seconds, every N samples, or on demand).
Minimum flight length before the first assignment. std from a few seconds of data is noisy.
Window: all samples since flight start, or a rolling window.
Where the assignment is shown (console, relay GUI, an iMotions event, a file).
When to refit the reference cohort as more pilots are logged.
Describe alternatives you've considered
Refit the whole cohort on every refresh. Rejected because KMeans numbers groups differently on each fit, so a pilot's group could change label without any change in how they fly. It would also move every other pilot's assignment.
Additional context
Blocked by #216 (batch fit to save), #211 (pilot_id/flight_id) and #213 (targets; without them there is no deviation to compute). Units must be post-#196, see docs/telemetry_schema.md.
One more open question: the client's current data is from Prepar3D at about 1 Hz, and the move to X-Plane (our plugin, 20 Hz) comes later. Group centres saved from a Prepar3D batch fit may not apply to live X-Plane flights. The flight models differ, and std depends on the sample rate. Before live assignment uses a Prepar3D reference cohort, check that the same flying gives the same features in both simulators, or resample to a common rate.
Answers from the client meeting (`docs/client_meeting_notes.md`):
Real-time clustering is wanted, and the client does not expect it to be computationally heavy.
It should run on the iMotions computer (the client already shows the pilots' views there), not on the simulator machine. That answers where the code runs. Where the result is shown is still open; we assume not to the pilot, since it could add stress.
Refresh rate is open. The client said it depends on the sensor rates: simulator about 90 Hz, EEG 256 Hz, EDA 128 Hz, eye tracking 51 Hz (as transcribed, unconfirmed). iMotions synchronizes timestamps. The client is not sure there is a common rate to cluster at.
The Prepar3D export is 1 Hz and the client cannot change it (closed software). That reinforces the earlier comment about a Prepar3D reference cohort not matching X-Plane at 20 Hz.
Status update on the blockers listed here. #216 merged in #225 and #211 merged in #219, so the saved batch fit and the pilot_id/flight_id columns both exist now. #247 adds the other missing piece, a live stream the Python side can read while iMotions is recording.
That leaves #213. Without targets there is no deviation to compute from real flights, so a live assignment can only be demonstrated on synthetic pilots. The open questions in this issue about refresh rate, minimum flight length and where the assignment is shown are still unanswered.
Activity
jmacd867 commentedon Oct 15, 2026
One more open question: the client's current data is from Prepar3D at about 1 Hz, and the move to X-Plane (our plugin, 20 Hz) comes later. Group centres saved from a Prepar3D batch fit may not apply to live X-Plane flights. The flight models differ, and
stddepends on the sample rate. Before live assignment uses a Prepar3D reference cohort, check that the same flying gives the same features in both simulators, or resample to a common rate.jmacd867 commentedon Oct 16, 2026
Answers from the client meeting (`docs/client_meeting_notes.md`):
Phlabry commentedon Oct 23, 2026
Status update on the blockers listed here. #216 merged in #225 and #211 merged in #219, so the saved batch fit and the
pilot_id/flight_idcolumns both exist now. #247 adds the other missing piece, a live stream the Python side can read while iMotions is recording.That leaves #213. Without targets there is no deviation to compute from real flights, so a live assignment can only be demonstrated on synthetic pilots. The open questions in this issue about refresh rate, minimum flight length and where the assignment is shown are still unanswered.