AIGP · PQ-FORECAST
Forecast based on VADR-TS-002 + A2RL × DCL precedent + AIGP confirmed

What the physical
qualifier will look like.

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.

Best-guess headline Synthesized
170±30 m
lap length
A2RL precedent
22±4
gates
A2RL had exactly 22
2laps
to validate
A2RL pattern · cited
~10m/s
avg speed
~150 km/h on straights
indoor · ~30×40 m floor · LED ceiling lighting · time-trial · all-gates-or-DQ
Confidence tags: CITED directly from primary source · footnoted ESTIMATE synthesized from precedents · flagged

01 · Direct precedent

A2RL × DCL Abu Dhabi · April 2025.

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.

Course geometry · cited
VenueADNEC Marina Hall, Abu Dhabi CITED
Indoor / outdoorIndoor hangar CITED
Track length170 m CITED
Gate count22 CITED
Laps for valid finish2 CITED
Winning time17.225 s · MavLab (TU Delft) CITED
Per-lap~8.5 s · ~10 m/s avg CITED
Top speed>150 km/h on straights CITED
Environment · cited
LightingIrregular, intentionally inconsistent CITED
Visual markersHardly any · "visually sparse" CITED
Drone computeNVIDIA Jetson Orin NX CITED
SensorsSingle forward camera + IMU CITED
ShutterRolling shutter CITED
Formats runGrand Challenge · AI-vs-Human · 4-up multi-drone · AI Drag CITED

Source: UAS Vision Apr 2025 · DCL $1M release

02 · What AIGP has confirmed

The cited facts.

From AIGP's own materials (theaigrandprix.com, Anduril press, official rules). Distinct from the A2RL precedent — these are AIGP-specific commitments.

DateSeptember 2026 · Southern California CITED
Indoor / outdoorIndoor · "consistent lighting for fairness" CITED
Obstacles"Expect obstacles and visual distractions" CITED
Drone vendorNeros Technologies · standard-issue identical hardware CITED
Compute~100 TOPS · single FPV camera ~12 MP wide-angle CITED
TelemetryIMU + motor RPM + battery + WiFi/BT CITED
NoLiDAR, GPS, depth sensor, global shutter CITED
ScoringTime-based · all gates required · missed gate = invalid run CITED
On-site block2-week training/qualification in SoCal pre-finals CITED
FinalsNovember 2026 · Columbus, OH (JobsOhio partnership) CITED
Finals distractionsSpectator cameras / flash possible CITED

Sources: theaigrandprix.com · Anduril press · Official rules

03 · Predicted course

Best-guess synthesis.

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.

Illustrative layout · indoor hangar ~30×40 m north up · scale 1:200
HANGAR · ~40 × 30 m floor · LED ceiling N ↑ SLALOM ZONE FAST STRAIGHT · 150 km/h HAIRPIN ELEVATION TRAVERSE FINISH SWEEP G1·start G22·finish judges + safety 5 m
start (1) standard (16) elevation / hairpin (3) traverse / drop (3) finish (1)
Geometry · estimate
Lap length170 ± 30 m EST
Gate count18–25 (A2RL=22) EST
Laps required2 EST
Average gate spacing7–10 m EST
Tightest sequence3 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
Venue · estimate
Floor area~30 × 40 m (1,200 m²) EST
Ceiling height10–14 m EST
LightingLED ceiling · ~500–1000 lux EST
Color temp4000 K cool-white EST
Floor surfacePolished concrete or matted EST
Walls / backdropBlack drape or matte gym walls EST
WindNone (HVAC drift <0.3 m/s) EST
CrowdNone — "no audience" per spec CITED

04 · Gate detail

Same gates the sim shows. Bigger than MultiGP standard.

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.

Gate · VADR-TS-002 §3.7 spec
Outer dimension2700 × 2700 mm · solid blue panel CITED
Inner aperture1500 × 1500 mm clear flight zone CITED
Border600 mm wide on each side CITED
Depth260 mm CITED
Corner bevel140 mm × 45° on the four outside corners CITED
ColorSolid 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
Detection budget

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.

Detect range (mAP>0.85)~80 m
PnP convergence range~40 m
Confident pose lock<25 m
Frame budget50 ms/frame

05 · What this means for training

Where to spend the next 4 months.

Implication 01
Indoor controlled lighting

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.

Implication 02
"Obstacles + distractions"

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.

Implication 03
2 laps to complete

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%.

Implication 04
All-or-nothing scoring

Miss one gate → run invalidated. Per AIGP rules. Conservative completion beats aggressive time-trial for the qualifier. Push speed AFTER VQ2 results, not before.

Implication 05
~10 m/s avg, 40 m/s peak

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.

Implication 06
Single forward cam · rolling shutter

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.

Implication 07
Tight indoor (~30 × 40 m)

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.

Implication 08
2-week on-site block pre-event

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.

Implication 09
100 TOPS compute · big upgrade

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

Most-likely subsection mix.

Subsection Gate count Length Tests Confidence
Start gate1arm/takeoff acquisitioncited
Fast straight3–5~25 mtop speed · acceleration · throttle envelopeest
Slalom alternating5–7~40 maggressive yaw · sub-3 m spacingest
Hairpin (180°)2–3~15 mdecel · bank-coordinated turn · re-acquisitionest
Elevation pair2~10 mvertical traversal · altitude PIDest
Sweeping curve3–4~30 msustained turn · steady yaw rate · bankingest
Vertical (rotated 90°) gate1unusual orientation · roll handlingest
Finish gate1crossing detection · run validationcited
TOTAL 18–24 ~170 m ~2× lap = ~340 m total flown

07 · Wild cards

Things that could surprise us.

Wild card 01
Multi-drone heats

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.

Wild card 02
Course revealed at site

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.

Wild card 03
Color shift

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.

Wild card 04
Sensor calibration mismatch

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.

Wild card 05
Reset / re-run policy

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.

Wild card 06
Standard drone, no mods

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.

Action items

Four months out.
Five things to get right.

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%.

01
Build hard-negative dataset
non-gate blue objects · banners · AV gear
May
02
Push cruise to 15–25 m/s
PID retune · throttle reserve · rolling shutter aug
Jun
03
Build VQ2 perception-aware policy
APEX PPO with detector + telemetry obs
Jul
04
Bring on-site retune kit
cal target · log analyzer · replacement weights
Aug
SoCal · 2-week on-site block
tune to real course · run to lock time
Sep 2026