Seoul traffic — experiment journal
Research record
- Question
- Can a marked virtual car find and drive a route through a bounded Seoul road network with background traffic?
- Evidence
- OpenStreetMap road extract with ODbL attribution; SUMO 1.28.0 OSM importer and microscopic traffic simulator. Actual trajectories and trip statistics linked below.
- Experiment
- Executed on Linux on 2026-10-28: three 1,800-second accelerated scenarios, 120/600 background trips, driving seed 27 and demand seed 20261028. Marked trip departs at 120 seconds. Fixed driver speed factor and noise; the two busy inputs differ only by marked-car rerouting.
- Results
- Marked durations 244.0 / 247.0 / 184.5 seconds; all 1,323 trips arrived across the three scenarios, zero reported collisions/teleports. A separate repeat matched complete parsed trajectories and route histories.
- Limitations
- Synthetic demand, inferred/default road attributes and signal timing. One demand seed, no statistical generalization, no live Seoul traffic, no learned autonomy or biological circuit controller.
- Reproduction
- Archived OSM/network, source hashes, environment lock, runner, full FCD, trip statistics, routes, converter warnings, logs and an independent repeat-run receipt are published. See the reproduction record.
Sources
Run evidence
Scope correction · 2026-10-28
This prototype uses SUMO driving rules and route finding. A fly neural circuit does not control the car, so it does not satisfy the brain-only requirement. It is preserved as map/traffic infrastructure, separate from the full-circuit feedback test. Brain-driven Seoul travel remains not run.
2026-10-28 · First drive
Open the recorded road experiment →
Task: Route a marked vehicle from the City Hall area to the Anguk area, while other vehicles share the roads. This is a separate traffic experiment; the learning bug does not drive the car.
Executed: Linux, SUMO 1.28.0, an archived OpenStreetMap extract, three accelerated 30-minute scenarios. The road geometry comes from OSM; demand, driver rules and signal plans are synthetic. Nothing connects to a physical vehicle.
| Scenario | Background trips | Marked trip | Waiting | Distance |
|---|---|---|---|---|
| Light, fixed route | 120 | 244.0 s | 50.5 s | 2.54 km |
| Busy, fixed route | 600 | 247.0 s | 47.5 s | 2.54 km |
| Busy, rerouting | 600 | 184.5 s | 23.0 s | 2.11 km |
The two busy cases share the same background trip elements and marked initial route, departure, seed and driver performance. Rerouting changes only the marked car’s device. In this scenario its trip shortened by 62.5 seconds. The modeled route costs and signal/queue interactions also matter; this does not isolate congestion as the only cause or prove a general routing advantage.
Verification: All 1,323 scheduled trips across the three cases arrived within the horizon. SUMO reported zero collisions and teleports. A separate Linux execution matched every parsed vehicle trajectory, speed, marked trip and route history. Legal route connectivity and source-to-replay checks run in CI.
Corrections retained: The first setup lacked a coordinate-conversion dependency. A preliminary run then exposed differing random speed factors; that comparison was excluded, and the published runs use speed factor 1.0, speed deviation 0 and car-following sigma 0. Original attempts remain preserved.
Next — not run: Test multiple demand/driver seeds, add a controlled blocked road, and compare rerouting responses. Rush-hour calibration, pedestrian interactions and a learned driver are not implemented.
Reproduce: Environment, commands, assumptions and full raw downloads. Run results, repeat check, file hashes.
Source record
OSM supplies road data under the ODbL; it does not supply measured traffic for this experiment. SUMO’s importer converts that data into a simulation network with defaults where needed. The exact extract and derived network are downloadable; no external map service is required during replay.
Update record
-
2026-10-28: Corrected the project boundary: this is supporting traffic infrastructure, not a fly-brain driving result. Removed it from current tasks; retained all original data.
-
2026-10-28: Downloaded road source, executed controlled Linux scenarios and a separate reproduction, published evidence and recorded replay. Preliminary failed/confounded attempts are documented; no real-world driving or learned policy was executed.