Comparing Human and Autonomous Drone Flight Behavior

Arm with throttle low, then throttle up. Fly the glowing gate in order — a light beacon marks the next gate and the page keeps per-track lap times. The tracks are the real gate layouts from the two datasets analyzed below: the TII RATM arena (4 gates, two straights with 180° turns and a 1.6→4.4 m altitude weave, gate headings taken from the human pilots' actual flight direction through each gate) and the UZH AlphaPilot figure-8 (10 gates, 50×20 m, flat and wave variants), coordinates lifted directly from the flight logs. Angle mode self-levels; Acro is real rate-mode flying with Betaflight-style expo — switch between them under Settings → Flight, along with the wind level.

RadioMaster / any RC transmitter: put the radio in USB joystick mode (EdgeTX: plug in USB → “USB Joystick (HID)”), then press a button or flip a switch once — browsers deliberately hide gamepads until a button is pressed, so moving the sticks alone won't wake it. (If nothing happens, map a switch to a button channel in the EdgeTX USB joystick setup.) It is read via the browser Gamepad API, so it works the same on this deployed site as it does locally. RadioMaster/EdgeTX devices use the browser layout observed for these radios. Firefox's standard-gamepad mapping exposes throttle as analog trigger/button 6 rather than an axis; the simulator handles that automatically. Other unmapped browsers fall back to the MoBI simulator's raw AETR channel order. No calibration is normally needed. If your model uses a different channel order, remap with Settings → Transmitter → Calibrate sticks (remembered only for that exact controller), or restore the built-in mapping with Use controller defaults. The arm switch is auto-detected: flip any spare switch once and that switch arms/disarms from then on. The input: readout in the control bar shows the live source the sim is receiving. With a radio connected, A arms, D disarms, and Space resets; after a crash, disarming respawns you. The Gamepad API needs a secure context — HTTPS, or localhost during development. Linux Firefox note: Firefox only exposes a gamepad after a button press, and RC radios in classic joystick mode emit none — use a Chromium-based browser to fly with a transmitter.

Hit ● REC and every armed span becomes a logged episode — 100 Hz state (position, velocity, attitude, body rates, motor speeds), your stick inputs, the motor commands the flight controller issued, and a per-step reward (gate progress + pass bonus − crash penalty − control effort, the same shape as the task reward in the research below) — plus webm video of the canvas. It streams as one NDJSON session to the destination you pick under Settings → Recording: browser download, or local disk / S3 via the same local bridge the EEG page uses (python -m eeg_lab.serve). Each episode also carries what a learner needs but a screen recording loses: the gate layout it was flown on, attitude and body rates converted into the research code's frame, and the rate setpoints the controller derived from your sticks — in Angle mode the sticks are tilt commands, so the setpoints are the only action channel that means the same thing in both modes. python -m ml.drone_rl.browser_data converts recordings into per-episode (obs, action, reward, done) datasets for offline RL, and python -m src.browser in ragtm/ converts them into dataset-format flights, so your own laps join the RATM pilots as demonstrations for the agents below.