A2RL RacingSimulator

A quadrotor simulator calibrated on real flight logs, used to plan and fly time-optimal laps through an A2RL-style gate course.

Year: 2026
Field: Autonomous drone racing
Tools: Python · PyTorch · CUDA
Status: Stage 1 of 4

Fig. 0 · S1 course, two laps T 00.00 sV 00.0 m/sALT 0.0 m GATE 00/22LAP 1/2

Fig. 0 — Replay of the simulated flight. Position and speed are the simulator's; attitude is smoothed for display (tilt ≤ 70°). The hall is a point cloud, revealed as the drone's forward camera (155° × 115°) scans it.

Problem

Autonomous drone racing asks a vehicle to fly a gate course faster than a human pilot, using only what it carries: one camera and an inertial sensor. In the A2RL format, a course has eleven gates, a split-S that dives through a gate directly below the previous one, and two laps against the clock.

Racing lines live at the edge of the airframe's envelope. A simulator is only useful if it is wrong in the same places the real drone is, so the model has to be fitted to real flights before any lap time means anything.

Model

Rigid-body dynamics with a Betaflight-style inner rate loop, rotor lag and a thrust model that grows with forward speed. Thousands of drones are simulated in parallel on the GPU.

Parameters were identified from 36 public flight logs of the A2RL airframe. On flights held out of the fit, predicted thrust reaches R² = 0.96 and the body-rate response R² = 0.93–0.95.

Rigid-body dynamics
System identification
PyTorch
CUDA

Fig. 1 · Racing line

03 · Course

Eleven gates, one hall.

A 100 × 30 m hall, square gates with a 1.5 m opening. The layout is reconstructed from published drawings; real coordinates were never released.

03.1 · Plan

A minimum-snap line through every gate.

Polynomial segments through the gates, with segment times optimised so the lap is as short as the motors allow. A tracking controller then flies it, compensating for drag.

03.2 · Speed

Fast on the straights, careful at the gates.

Colour shows speed along the flown line. The drone peaks on the long straight and brakes hard before the turn-around.

03.3 · Split-S

Through the top gate, back through the one below.

The hardest manoeuvre on the course: a half-loop that reverses heading and drops a gate height in under a second.

04 · Results — S1 course, two laps

18.63sRace time, first gate to finish

25.9m/sTop speed

7.7gPeak acceleration

Fig. 2 · Speed

Fig. 2 — Speed over the race. Thin lines mark gate passages; the dashed line starts lap two.

04.1

Against the real race

MAVLAB, TU Delft · 1st16.56 s
This simulator18.63 s
Human world champions19.10 s
KAIST · 3rd, final36.80 s

Real results: A2RL, April 2025. Simulated course is approximate (±2–3 m).

Fig. 3 · Speed vs. reliability

Fig. 3 — 256 drones per plan, each with perturbed mass (±5%), max thrust (±10%), drag (±30%), motor lag (±30%), inertia (±20%) and speed-dependent lift (±30%). Median times. Faster plans finish less often.

Next: knowing where it is

So far the simulator flies on perfect state. The next stage replaces it with what the real drone sees: a wide-angle camera, gate corners detected in the image, a pose from PnP and an extended Kalman filter fusing it with the IMU at 17 ms latency.

Fig. 4 · Kalman filter Running in your browser

Fig. 4 — The IMU predicts, noisy camera fixes correct, the ellipse is the 2σ uncertainty. It grows whenever the camera drops out.

Limits

01
Course coordinates are reconstructed from drawings, accurate to roughly ±2–3 m.
02
Mass and inertia were not measured on the real airframe.
03
Logs cover speeds up to 21 m/s; above that the thrust model is an extrapolation.

Next02 / 03 VARIGEN →