Two-track plan: a deterministic VQ1 completion stack (YOLO + PnP + PID) ships first to lock in a passing submission. APEX perception-aware PPO takes over for VQ2 speed. No GPS. No absolute positioning. FPV + telemetry in, Throttle / Roll / Pitch / Yaw out. Per the confirmed 2026-04-19 AIGP spec.
Same input surface and same output surface for both tracks. Only the controller changes.
The opinionated master strategy. Effort budget across stages, reliability math, data pipeline moat, sim-to-real bridge, compute envelope, explicit anti-patterns, scorecard.
Tactical expansion of the playbook. VQ1 vs VQ2 stacks, training pipeline, detector choice, retired components, Sim Day 1 checklist, intel.
Three-phase training — YOLO11n detector → YOLO11n-pose keypoints → perception-aware PPO. Phase 3 observation-swap flag required for VQ2 transfer.
What changed in the 2026-05-08 spec revision (camera intrinsics, UDP vision stream, ODOMETRY removed) and how the code was aligned on 2026-05-12.
VADR-TS-002 compliance checklist, 3-tier training strategy, submission pipeline, risk matrix.
End-to-end pipeline diagram, component file map, sensor budget, commands surface.
Six states, transition rules, latency budget for a single frame-to-command cycle.
How the APEX detector chain + telemetry adapter + controller bolt into the sim client.
Detector choices (YOLO11n, RF-DETR), dataset layout, evaluation harness, confidence thresholds.
Local Windows / FastAPI dashboard. Live GPU telemetry, one-click controls, run history, live log tail. Double-click launch_trainer_app.bat.
Exact training commands + the overnight_autotrainer.py CLI. Unattended nightly runs with precheck, backup, benchmark, auto-promotion, rollback.
Two-track plan: VQ1 completion stack + VQ2 APEX PPO with observation swap. Parallel sim instances, dataset expansion for distractors.
Auto-label training data by screen-capturing DCL The Game while you play. Closest VQ2-realistic imagery we can get pre-sim. ~3-5K frames per 2-4 hr session.
Feed the detector gate-lookalike images (arches, fans, scaffolding) with empty labels to suppress VQ2 false positives. Keyword-based harvest + auto false-positive discovery.
Three-tier plan to close the "nothing handles obstacles" gap the AIGP spec exposes. Headless Playwright + race-r3f.html produces auto-labeled nc=2 training data. Free synthetic imagery.
One-command APEX run on the RTX 5080 box (~7.5 hr overnight). Outputs, weights, troubleshooting.
Detector benchmark harness — YOLO11n vs RF-DETR vs U-Net on the same course proxy.
Gains, thresholds, timeouts. One source of truth for PID + PPO hyperparameters.
Windows-only sim, anti-cheat internet, T/R/P/Y outputs, nine-check submission list, portal TBD.
Clone, install deps, smoke test, first training run. For contributors joining the repo fresh.
Minimum viable run: detector weights in, frame in, command out. Single-file reproduction.
CUDA OOM, observation-mode confusion, submission-validator failures, anti-cheat handshakes.
Full BOM, frame, FCU, ESCs, motors, cameras. Practice-rig recipe for physical-round prep.
Printable carrier shell for the Neros Archer platform (supplied at physical qualifier).
Mounts, camera brackets, antenna holders — everything printable for the build.
Transmitter + receiver combos validated for development. Protocol + latency notes.
Assembly steps, soldering order, first-flight checklist. Updated for the 2026 practice rig.
Dimensioned drawings for frame + camera mount. Source of truth for BOM disputes.
How we win. Effort budget, reliability math, data pipeline moat, sim-to-real, anti-patterns, scorecard.
Deployed dashboard URLs, Cloudflare Pages, Workers, upstream AIGP resources.
Phase-by-phase details: detector, keypoints, perception-aware PPO. Observation schemas for both modes.