Compliance checklist mapped to AIGP rules (VADR-TS-002, supersedes -001 · 2026-04-19 update), submission pipeline, risk matrix, 8-week timeline. Every row is a "must be true" gate for a valid submission — not a nice-to-have.
python submit_check.py package| # | Requirement | Source | Status |
|---|---|---|---|
| 1 | Code in Python | Rules §1 | OK |
| 2 | No human interaction during runs | Rules §3 · 2026-04-19 | OK · deterministic + learned policies only |
| 3 | Runs on Windows | 2026-04-19 update | OK · smoke on Win11 |
| 4 | Active internet during runs (anti-cheat) | 2026-04-19 update | OK · no offline-only deps |
| 5 | No GPS · no absolute positioning usage | 2026-04-19 update | OK · audit: no NED in runtime code |
| 6 | No depth sensor usage | 2026-04-19 update | OK |
| 7 | Outputs are Throttle/Roll/Pitch/Yaw | 2026-04-19 update | OK · ControlCmd schema matches |
| 8 | Run ≤ 8 min | Rules §5 | OK · race.max_time_s = 480 |
| 9 | No input() / keyboard reads in runtime | Rules §3 | OK · audited |
| 10 | Code reviewable (pinned deps, reproducible) | Rules §7 | PARTIAL · need reproduce.sh |
| 11 | No third-party services called at runtime | Rules §7 | OK · local inference only |
| 12 | Weights ship with submission | Rules §2 | PARTIAL · packager needs --include-weights |
| 13 | No imitation learning from human demos | Rules §3 (inferred) | OK |
| 14 | Detector handles <10 gates (VQ1) | 2026-04-19 update | OK |
| 15 | Detector handles <20 gates + distractors (VQ2) | 2026-04-19 update | OPEN · needs sim-frame fine-tune |
| 16 | PPO observation matches real sim surface | 2026-04-19 update | PARTIAL · stub in place, needs real YOLO wiring |
| 17 | Gate ordering without privileged map | 2026-04-19 update | OK · target-gate tracker |
| 18 | No team-shared accounts (one team per participant) | 2026-04-19 update | OK · organizational |
| 19 | No full-time founding-partner employees | 2026-04-19 update | OK · confirm team roster |
| 20 | Multiple parallel sim instances work in training | 2026-04-19 update | OPEN · SubprocVecEnv wrapper |
| 21 | On-drone compute fits Jetson Orin NX class (Physical) | Rules §6 | OPEN · TensorRT INT8 export |
| Tier | Scope | Dataset | Output |
|---|---|---|---|
| Tier 1 · Pre-sim | APEX Phase 1 + 2 on existing data. VQ1 pilot ready. | dataset_gates_mega (2.7K) + hard negatives (1K) + synthetic (future) | Submission-ready detector for Day 1 of VQ1 |
| Tier 2 · Post-VQ1 | Fine-tune on captured sim frames. Train PPO with detector_telemetry obs. | +6K VQ1-sim frames + continuous capture | PPO policy transferable to real AIGP sim |
| Tier 3 · Sim-to-real | Residual dynamics model from real flight data. Adapt detector to real imagery. | ~30 min DIY rig flight · real FPV | Policy that transfers to Physical Qualifier |
| Risk | Likelihood | Impact | Mitigation |
|---|---|---|---|
| Privileged-obs PPO in submission | Medium | HIGH | --observation-mode=detector_telemetry default · submit_check warns |
| Detector overfits synthetic data, fails on real sim | High | HIGH | Sim-day capture + fine-tune on Day 1 · continuous capture loop |
| Anti-cheat handshake breaks training | Medium | MEDIUM | Test parallel-instance handshake in first sim access |
| Sim-to-real fails at Physical | Medium | HIGH | 30 min real-flight data + residual dynamics in Aug · DIY rig testing |
| Compute over budget on Orin NX | Low | MEDIUM | INT8 TRT export · budget verified at 10 ms · P2 pruning pass |
| Rule change mid-competition | Medium | MEDIUM | Modular stack · component replacement <1 day · monitor forum |
| Team-eligibility audit surprise | Low | CRIT | Confirm every participant once · single-team rule understood |
| PPO policy non-deterministic at submit time | Low | MEDIUM | Deterministic inference flag · seed fixed · smoke on 100 repeats |
| Week | Focus | Deliverable |
|---|---|---|
| W1 (now) | Theme refresh · strategy docs up · sim package watch | Docs complete · VQ1 pilot smoke-tested |
| W2–3 | VQ1 sim access · telemetry adapter · first completion | 100% gate-clear on VQ1 course |
| W3–4 | Sim-frame capture · detector fine-tune · submit VQ1 | VQ1 submission locked |
| W4–5 | PPO observation swap · parallel envs · Phase 3 training | PPO policy beats PID on VQ1 course |
| W5–6 | VQ2 opens · PPO tuning · adversarial detector | Multiple scored VQ2 attempts |
| W6–7 | VQ2 submissions · chaos testing · hero run | Top-30 seed for Physical |
| W7–8 | VQ2 cutoff · DIY rig real-flight capture | Real-flight dataset >15 min |
Master strategy: effort budget, reliability math, sim-to-real, data pipeline moat, anti-patterns.
Three-phase training · observation-swap details · research foundations.
Validation, benchmark, package, 9-check submission list.
End-to-end pipeline, component file map, comms surface.