AI GRAND PRIX // RACER.DEV
Autonomous Racing · Practice Build & Autonomy Stack

You Don't
Build the Drone.
You Build the Mind.

The AI Grand Prix presented by Anduril is a spec-airframe, zero-mod, software-only race — every team flies an identical Neros drone, so the only thing you bring is autonomy code. This is your max-spec practice rig + autonomy stack to develop and tune that code before sim-to-real.

Prize Pool
0K + a job
Teams in 24 h
0+
Live Final
NOV 2026 · OHIO
You Build
CODE · not hw
// Seedance 2.0 — generated flight reelHQ · 720p · onboard view
AUTONOMOUS // NO HUMAN INPUT VEL 142 KM/H INFERENCE Δ 6.2 ms · 157 TOPS SEEDANCE 2.0 · REC
// machine-vision overlay — how the autonomy sees itvision-only · 157 tops
GATE 0.71 GATE · 0.98 AUTONOMOUS // NO HUMAN INPUT VEL 142 KM/H INFERENCE Δ 6.2 ms GATE 03 / 12 · LAP 2
01

The Competition Brief

Announced 27 Jan 2026 · Palmer Luckey
// read this first

The hardware is settled. Now, show us the code.

Every team races an identical drone built by Neros Technologies — hardware modifications are prohibited. Performance comes entirely from each team's onboard autonomy software. You will not build the race drone. The parts on this page are for a practice/dev rig that mirrors the likely platform so you can build and test your stack before the real drones are in your hands.

"A pure-play test of computer vision, path planning, and edge computing." — Anduril
who

The Players

Anduril hosts · Neros builds the spec drones · Drone Champions League runs race ops + the simulator (its "AI vector module" is the drone's brain) · JobsOhio hosts the finals.

prize

What You Win

$500K pool split across top finishers — and the headline prize is a job at Anduril (top scorer skips the queue to a hiring-manager interview; US security clearance needed for the offer).

rules

The Format

Fully autonomous, zero human control — any piloting/assist over a link = DQ. All compute onboard. Teams of up to 8. Open to universities + independent engineers worldwide.

// Timeline — the simulator is the gate, not the hardware

01
Spring 2026
Remote Qualifier

Submit Python autonomy code to race inside the DCL simulator. This is the entry gate — no hardware required.

02
Sept 2026
Sim-to-Real

Top teams flown to Southern California for a ~2-week round adapting their code onto the real Neros drones.

03
Nov 2026
Live Championship

Head-to-head autonomous racing in Columbus, Ohio, near Anduril's Arsenal-1 campus.

Register
theaigrandprix.com

Sign up, get sim access, start writing code. That's the real work.

02

Brain ≠ Pilot

Two devices · two jobs · not substitutes

The confusion to kill up front: a Jetson carrier and a flight controller are not interchangeable. An autonomous drone needs both. The Jetson is the brain; the flight controller is the pilot. The autonomy software runs on the Jetson module — and it's identical no matter whose carrier board hosts it (DAMIAO, DFRobot, Holybro). The carrier brand changes nothing about your code.

// the brain

Jetson Orin NX

on a DAMIAO carrier

Runs NVIDIA JetPack (Linux) + your perception & planning stack. Sees the gates, estimates state, decides the trajectory. This is what you actually compete with. Carrier-agnostic software.

// the pilot

Flight Controller

light Betaflight rate FC

Runs the low-level rate loop (gyro → motor mixing) at kHz and handles failsafe. Takes collective-thrust + body-rate commands from the brain. Not the heavy 6X Pro — wrong tool for a 5-inch.

The proven pattern (UZH Agilicious, A2RL): the Jetson does all the smart work and streams collective-thrust + body-rate setpoints to a few-gram Betaflight FC that closes the fast inner loop. Keep your Pixhawk 6X Pro for the cinema rig or a larger airframe — its weight and triple-redundant sensors are an asset there, a penalty here.

03

Max-Spec Practice Rig

Mirrors the likely race envelope

Sized to the strongest available proxy — the A2RL championship, run by the same operator (DCL) whose "AI vector module" is in the Neros drone: Jetson Orin NX compute, single forward RGB camera + IMU, vision-only, fully onboard, small high-speed quad. Maxed for headroom; tune down once Anduril publishes real specs.

SystemMax-Spec PickWhy≈ Price
ComputeAI brainNVIDIA Jetson Orin NX 16GB Super
157 TOPS · up to 40W · 260-pin SO-DIMM
Top of the SO-DIMM class — memory headroom for vision nets + VIO. Same family A2RL races. Run MAXN Super, cool it well.~$700–970
CarrierJetson baseboardDAMIAO Orin NX Carrier V1.1
36 g · 87×50 mm · GbE/USB3/CSI/M.2/XT30
Featherweight Jetson-only carrier (vs 203 g Holybro). NOT a flight controller — you add the FC separately. Verify CSI/M.2 before buying.~$30–135
StorageLogging / datasetsNVMe M.2 SSD 512 GB
2242/2280 PCIe
ROS 2 bags, training clips, model store. Onboard logging is gold for sim-to-real debugging.~$45–70
CoolingThermalActive heatsink + fan
for MAXN Super 40W
Super mode draws real watts; static/hover hot-soak is worst case. Unobstructed airflow mandatory.~$20–40
Flight CtrlThe pilotLight Betaflight AIO (H7) + 4-in-1 ESC
30.5×30.5 stack · ~10–15 g
kHz rate loop + failsafe; UART/CRSF/MSP bridge to the Jetson. The Agilicious/A2RL inner-loop pattern. Not the 6X Pro.~$80–140
SensorVisionGlobal-shutter RGB cam (CSI) + FC IMU
e.g. Arducam/IMX-class global shutter
Global shutter kills motion skew at race speed; single forward cam + IMU = the vision-only proxy. Avoid GPS/LiDAR dependence.~$60–250
AirframePractice quad5-inch carbon race/freestyle frame
~210–240 mm
Cheap, agile dev platform. The race drone is sub-8-inch; a 5-inch is a reasonable stand-in for autonomy work.~$40–90
DrivetrainMotors / props~2207 1960–2700 KV · 5" tri/bi-blade
6S class
Standard 5-inch power. Doesn't affect autonomy work; pick for thrust + flight time on the dev bench.~$60–110
PowerBattery6S LiPo 1300–1800 mAh ×several
+ powers the Jetson via XT30
Rotate packs for long dev sessions. Confirm carrier input range covers your pack voltage.~$25–40 ea
SafetyManual overrideELRS RX + radio (+ optional FPV)
human kill-switch / takeover
Essential during dev — a safety pilot must be able to seize control and disarm. Non-negotiable for testing autonomy.~$60–250

Official specs aren't public yet. Anduril says airframe size/weight, the exact compute model, and sensor rules come "at a later stage" (the rules page currently 504s). Everything above is an informed proxy from A2RL/AlphaPilot precedent — build modular so you can re-tune the moment real specs drop, and remember none of it is your competition entry.

Fig · Why These Picks — Compute & Carrier Weight

Jetson Orin — AI Compute

TOPS · sparse INT8 · Super mode

Orin Nano 8GB67
Orin NX 8GB117
Orin NX 16GB ★157

The 16GB gives memory headroom for vision nets + VIO at once — the pick for a max-spec rig.

Carrier Weight

bare board · grams

DAMIAO carrier ✓36 g
Holybro baseboard203 g

5.6× lighter. The Holybro bundles a Pixhawk FC; the featherweight DAMIAO is Jetson-only — right for a racer where you pair a separate few-gram FC.

04

The Autonomy Stack

Where the race is actually won

A starting architecture — from the metal up to the policy. Perception sees the gates, estimation tracks where you are, planning picks the line, control flies it. Train in sim, port to real with domain randomization (the UZH "Swift" playbook that beat human champions).

Data Flow · Camera → Motors
ON-BOARD JETSON ORIN NX · autonomy stack SENSORSRGB cam + IMU PERCEPTIONgate detectYOLO→TensorRT ESTIMATIONVIO · GPS-freecuVSLAM PLANNINGtrajectorymin-snap / RL CONTROLCTBR setpointthrust+bodyrate FLIGHT CTRLBetaflight rate loop→ ESCs / motors state / VIO pose feedback
PlatformOS · SDK
NVIDIA JetPack 6.2 on the Orin NX (Ubuntu 22.04, CUDA, cuDNN, TensorRT). Enable MAXN Super for the full 157 TOPS.
JetPack 6.2CUDA / TensorRTUbuntu 22.04
Middlewareglue
ROS 2 (Humble/Jazzy) for nodes + message passing; bridge to the flight controller over serial (MSP/CRSF) or MAVLink if you run a small PX4 board instead.
ROS 2MSP / CRSF bridgeuXRCE-DDS (if PX4)
Perceptionsee the gates
Gate detection — a YOLO-class CNN trained on race gates, exported ONNX → TensorRT for real-time edge inference. Optional monocular depth.
YOLO / detectorPyTorch → ONNX → TensorRTcorner regression
Estimationwhere am I
Visual-Inertial Odometry fusing the RGB cam + IMU — GPS-free state estimation. Start with a proven stack, accelerate on the GPU.
Isaac ROS cuVSLAMVINS-FusionOpenVINS
Planningpick the line
Gate-relative trajectory (minimum-snap) for a classical baseline, or a learned policy (deep RL trained in sim) for the Swift-style frontier. Hybrid is common.
min-snapdeep RL policyMPC (optional)
Controlfly it
High-rate controller emits collective-thrust + body-rate (CTBR) → the Betaflight FC closes the inner rate loop at kHz. This split is the autonomous-racing standard.
CTBR setpointsBetaflight rate loopfailsafe / arming
Simulationthe real gate
The DCL simulator is the Spring-2026 qualifier. Practice + train in open sims first, then port. Use domain randomization for sim-to-real transfer.
DCL sim (entry)Flightmare / AgiliciousIsaac Simdomain randomization
05

The Roadmap

80% software · 20% hardware

Execution Order

  1. Register at theaigrandprix.com; get DCL simulator access.
  2. Stand up the toolchain — JetPack 6.2, ROS 2, PyTorch, TensorRT.
  3. Build a gate detector + baseline min-snap planner in sim; get a clean lap.
  4. Assemble the practice rig (Orin NX + DAMIAO + Betaflight FC + global-shutter cam).
  5. Bring up VIO on real hardware; validate state estimate against motion.
  6. Close the CTBR loop Jetson → Betaflight; fly hand-tuned autonomous laps with a safety pilot.
  7. Iterate the policy in sim (RL + domain randomization), redeploy, measure lap time.
  8. Submit to the Spring-2026 remote qualifier — the actual gate.
  9. Sim-to-real prep for the Sept SoCal round on Neros hardware; then Columbus.

Where to Spend Effort

80%

Software & Sim

Perception, planning, control, and DCL-sim lap time. This is the entire competition — and the only thing that ports to the spec drone.

20% Hardware

The practice rig is a learning bench to prove your stack on real silicon and sensors. Keep it cheap, modular, and disposable.

The unfair advantage: the headline prize is a job — so polished, well-documented, reproducible autonomy code that a hiring manager can read is worth as much as raw lap time. Build like you're submitting a portfolio.