The April 2025 A2RL × DCL Autonomous Drone Championship in Abu Dhabi is the direct precedent: same sim vendor, same drones, same format. 170 m track, 22 gates, 2 laps required, indoor hangar. AIGP has confirmed Southern California indoors with consistent lighting and obstacles. This doc threads the cited facts into a single forecast you can train against.
01 · Direct precedent
Same sim vendor (DCL Simulator), same drones (Neros / Jetson-class compute), same gate spec, autonomous AI format. The closest reference point we have for what AIGP will actually look like in physical form.
| Venue | ADNEC Marina Hall, Abu Dhabi CITED |
| Indoor / outdoor | Indoor hangar CITED |
| Track length | 170 m CITED |
| Gate count | 22 CITED |
| Laps for valid finish | 2 CITED |
| Winning time | 17.225 s · MavLab (TU Delft) CITED |
| Per-lap | ~8.5 s · ~10 m/s avg CITED |
| Top speed | >150 km/h on straights CITED |
| Lighting | Irregular, intentionally inconsistent CITED |
| Visual markers | Hardly any · "visually sparse" CITED |
| Drone compute | NVIDIA Jetson Orin NX CITED |
| Sensors | Single forward camera + IMU CITED |
| Shutter | Rolling shutter CITED |
| Formats run | Grand Challenge · AI-vs-Human · 4-up multi-drone · AI Drag CITED |
Source: UAS Vision Apr 2025 · DCL $1M release
02 · What AIGP has confirmed
From AIGP's own materials (theaigrandprix.com, Anduril press, official rules). Distinct from the A2RL precedent — these are AIGP-specific commitments.
| Date | September 2026 · Southern California CITED |
| Indoor / outdoor | Indoor · "consistent lighting for fairness" CITED |
| Obstacles | "Expect obstacles and visual distractions" CITED |
| Drone vendor | Neros Technologies · standard-issue identical hardware CITED |
| Compute | ~100 TOPS · single FPV camera ~12 MP wide-angle CITED |
| Telemetry | IMU + motor RPM + battery + WiFi/BT CITED |
| No | LiDAR, GPS, depth sensor, global shutter CITED |
| Scoring | Time-based · all gates required · missed gate = invalid run CITED |
| On-site block | 2-week training/qualification in SoCal pre-finals CITED |
| Finals | November 2026 · Columbus, OH (JobsOhio partnership) CITED |
| Finals distractions | Spectator cameras / flash possible CITED |
Sources: theaigrandprix.com · Anduril press · Official rules
03 · Predicted course
Combining the A2RL precedent's hard numbers with AIGP's published constraints. Everything below is flagged ESTIMATE — these are the most likely values, not confirmed.
| Lap length | 170 ± 30 m EST |
| Gate count | 18–25 (A2RL=22) EST |
| Laps required | 2 EST |
| Average gate spacing | 7–10 m EST |
| Tightest sequence | 3 m gate-to-gate slalom EST |
| Longest straight | ~20 m fast section EST |
| Avg speed | ~10 m/s · ~22 mph CITED |
| Top speed straights | ~40 m/s · ~90 mph CITED |
| Floor area | ~30 × 40 m (1,200 m²) EST |
| Ceiling height | 10–14 m EST |
| Lighting | LED ceiling · ~500–1000 lux EST |
| Color temp | 4000 K cool-white EST |
| Floor surface | Polished concrete or matted EST |
| Walls / backdrop | Black drape or matte gym walls EST |
| Wind | None (HVAC drift <0.3 m/s) EST |
| Crowd | None — "no audience" per spec CITED |
04 · Gate detail
VADR-TS-002 §3.7 locks the gate spec for both VQ1/VQ2 and the physical qualifier — the rendered training assets are dimensionally identical to the physical panels.
| Outer dimension | 2700 × 2700 mm · solid blue panel CITED |
| Inner aperture | 1500 × 1500 mm clear flight zone CITED |
| Border | 600 mm wide on each side CITED |
| Depth | 260 mm CITED |
| Corner bevel | 140 mm × 45° on the four outside corners CITED |
| Color | Solid blue (PMS 286-ish from spec render) CITED |
| Vs MultiGP standard (5') | ~98% of standard span · identical scale EST |
| Vs MultiGP Champ (7') | 70% of Champ span EST |
A 1.5 m gate at 50 m subtends ~1.7° horizontal at fx=320 — about 9.5 pixels wide in the 640×360 frame. At 100 m it's ~4.8 pixels. At 25 m it's ~19 pixels.
05 · What this means for training
AIGP commits to "consistent lighting for fairness." This is easier than A2RL's irregular lighting. Our synthetic dataset trained on bright/consistent gates should transfer well. Don't over-invest in shadow/highlight augmentation.
AIGP-confirmed. Need a hard-negative training set with non-gate blue objects (banners, signage, AV gear). Detector must reject false positives, not just maximize recall.
A2RL pattern. Means our stack must reliably handle ~44 consecutive gates (22 × 2) without a single miss. Per-gate reliability target: 99.5%+ for cumulative > 80%.
Miss one gate → run invalidated. Per AIGP rules. Conservative completion beats aggressive time-trial for the qualifier. Push speed AFTER VQ2 results, not before.
Real speeds. Our VQ1 stack at 6 m/s target needs to step up. PID gain tuning at 15–25 m/s cruise, with throttle reserve for sub-second hairpin recovery.
A2RL hardware. No depth, no stereo, no global shutter. Rolling-shutter artifacts at 40 m/s = mandatory robustness in the perception loop. Add motion-blur + skew augmentation to synth data.
Estimated venue size. Gate-to-gate as short as 3 m on slalom. Detector handoff needs to fire on the same frame the previous gate exits the FoV — there is no acquisition gap.
AIGP-confirmed. Two weeks in SoCal on the real course. Bring instrumentation for in-situ retuning: scope, log analyzer, replacement weights. Don't assume you'll re-train from scratch.
Neros drones ship with ~100 TOPS (Orin NX class) — not the 67 TOPS Orin Nano we assumed in JET-ARCHER. Headroom to run YOLOv11s or even RF-DETR, not just YOLOv11n.
06 · Course-shape forecast
| Subsection | Gate count | Length | Tests | Confidence |
|---|---|---|---|---|
| Start gate | 1 | — | arm/takeoff acquisition | cited |
| Fast straight | 3–5 | ~25 m | top speed · acceleration · throttle envelope | est |
| Slalom alternating | 5–7 | ~40 m | aggressive yaw · sub-3 m spacing | est |
| Hairpin (180°) | 2–3 | ~15 m | decel · bank-coordinated turn · re-acquisition | est |
| Elevation pair | 2 | ~10 m | vertical traversal · altitude PID | est |
| Sweeping curve | 3–4 | ~30 m | sustained turn · steady yaw rate · banking | est |
| Vertical (rotated 90°) gate | 1 | — | unusual orientation · roll handling | est |
| Finish gate | 1 | — | crossing detection · run validation | cited |
| TOTAL | 18–24 | ~170 m | ~2× lap = ~340 m total flown |
07 · Wild cards
A2RL ran 4-up multi-drone with collision-avoidance scoring. AIGP rules currently describe time-trial, but the FAQ leaves room. Our stack has no collision avoidance — single-drone assumption baked in.
If AIGP releases the course geometry only at the on-site block (2 weeks pre-event), we won't be able to train a course-specific policy. Our stack must generalize from VQ2 sim to unseen physical layouts.
VADR-TS-002 renders gates in solid blue. If the physical paint is a different shade than the sim (sim RGB vs PMS 286 real), detector trained on sim color could misfire on real gates. Calibrate on-site, plan retraining time.
VADR-TS-002 fixes camera intrinsics (fx=fy=320, cx=320, cy=180, tilt +20°). Real cam may drift by ±5%. PnP estimates degrade gracefully but altitude/lateral targets shift. Bring a calibration target.
AIGP says unlimited runs during qualification window. But are runs back-to-back or do we wait between? Each crash burns time but also props/batteries. Bring redundant hardware.
Drones are AIGP-issued Neros, identical hardware. No carrier PCB swap, no JET-ARCHER on race day. Everything is software. The hardware briefs are useful for post-AIGP, not for the qualifier itself.
The hardware is fixed (Neros / Orin NX). The gates are fixed (1.5 m inner). The venue is fixed (SoCal indoor). Everything we control is software. Train against the cited specs, build hard-negative datasets, push speed only after reliability hits 99%.