How we win the
AI Grand Prix.

§ 01What actually wins

The AI Grand Prix looks like a four-stage race: VQ1 → VQ2 → Physical → Final. It isn't. It's a filter with a prize at the end:

§ 02Reliability math > speed math

Each VQ2 crash costs roughly 15 minutes: the failed 8-minute run + restart + warm-up + anti-cheat rehandshake. In expected-time terms:

E[lap_time] = p(clear) · best_time + p(crash) · 15 min

A 45s pipeline with 5% crash → E = 0.95·45 + 0.05·900 =  87.75s equivalent
A 50s pipeline with 0.5% crash → E = 0.995·50 + 0.005·900 = 54.25s equivalent

                   slower pipeline wins by 33s per attempt

How to operationalize this:

• Run identical conditions 10+ times before tuning any parameter, so you can tell signal from noise.
• Log every attempt. A single crash at run 87 in an otherwise-clean regime = investigate, don't ignore.
• Penalty weighting in PPO reward: r_crash = -500, not -50. Make the policy genuinely crash-averse.
• Fallback path when detection drops: slow forward + gentle yaw sweep (not "hold attitude and pray"). We already do this in the VQ1 pilot; port the logic into the PPO env.

§ 03The data pipeline is the moat

Unlimited attempts × 8-min runs × 10 fps camera = 4,800 frames per attempt. By the VQ2 deadline, a team that captures and auto-labels every attempt will have 500K+ in-distribution training frames. A team that doesn't will still be training on the initial 2,000 synthetic frames.

The plumbing that unlocks this:

1. Record everything. Frame + telemetry + detection + our controller output + outcome, every run, every attempt. Zstd-compressed, ~500 MB per attempt.
2. Auto-label with the current detector. Run Phase-1 YOLO on every captured frame overnight. Keep detections with conf > 0.7 as labels.
3. Human-correct the 5% edge cases. Hard negatives (false positives), low-confidence hits, frames where the pipeline crashed. 10 min a day, someone on the team.
4. Retrain weekly. Fine-tune detector on the growing dataset. Track mAP@50 on a held-out validation slice every week.
5. Feed the improved detector back into Phase 3 PPO training. Observation quality improves with the detector; policy gets stronger in lockstep.

§ 04Sim-to-real transfer is the hidden gate

Between VQ2 close (~late July) and Physical Qualifier (early September) is a ~6-week window. Most teams will relax here — bask in qualifier rankings, maybe tune one or two params. This is when the Physical is actually won.

What happens in the September gap

Why modular architecture was the right bet for this

Our stack is detect → keypoints → PnP → controller. Every boundary is swappable. At the sim-to-real step we keep the detector + keypoints (they transfer well because game-engine texture is already in-distribution for real-world-trained backbones), we update PnP intrinsics to the real camera, and we insert the residual-dynamics model in front of the controller. That's three files touched, not a full retrain. End-to-end pixel-to-motor policies don't have these seams — they retrain or they lose.

§ 05Using the unlimited-attempt budget

Most teams will treat submissions as hero runs — maybe 5 total, chasing a personal best. That wastes the single biggest experimental asset we have. Our plan:

§ 06Compute envelope (on-drone)

Physical Qualifier uses the Neros Archer platform. Compute spec published closer to the event, but based on class (100 TOPS @ ~15 W) assume Jetson Orin NX-class. Design for it now:

Target 10 ms gives us 2× headroom before the 20 ms per-frame budget (matches 50 FPS camera). Anything more is wasted on this hardware class.

§ 07Explicit anti-patterns (don't build these)

§ 08Calendar (from now → Ohio)

§ 09Team discipline — what we commit to

• No hero work. Every change goes through a reproducible command (train_apex.py phase, captured config, logged results). No "I just tweaked it in a notebook."
• Every attempt logged. JSONL per-frame dump. Zstd-compressed. Never deleted.
• Weekly retrain cadence. Friday evening: fine-tune detector on the week's captures, push, evaluate Sunday.
• One canonical repo. No forks, no side experiments in private. The submission always builds from main.
• Reviewable submission. Anduril can ask for the code. Pin dep versions, include reproduce.sh, structure the repo so a reviewer can trace any output back to deterministic inputs.
• Abandon-early rule. If a direction hasn't produced a signal in 5 days of focused work, stop. Reallocate to the fallback. (This is how we killed LingBot-Map in 48 hours.)

§ 10Scorecard (how we know we're winning)