vdmaas98/uzh-fpv-drone-racing-128

This dataset was created using LeRobot.

UZH-FPV Drone Racing, 128px, LeRobot format

A processed derivative of the UZH-FPV Drone Racing Dataset packaged as a LeRobotDataset (v3): 24 episodes, 24,242 frames of aggressive indoor FPV racing flight, as 128x128 grayscale onboard camera frames with 4D continuous control actions.

This is not my recording. All flight data was collected and published by the Robotics and Perception Group, University of Zurich. This repository redistributes a processed version under the same CC BY-NC-SA 3.0 license, with attribution and modifications listed below. It does not imply any endorsement by UZH.

What was changed relative to the original

• Camera frames from the indoor sequences (both the forward-facing and the 45-degree downward-tilted camera tracks, left and right fisheye cameras) were converted to grayscale and downsampled to 128x128.
• Frames were timestamp-aligned to the ground truth and IMU streams.
• Actions are DERIVED, not recorded pilot inputs: attitude from the ground-truth quaternions and IMU was converted to equivalent ANGLE-mode stick positions (angle limit 60 degrees, max yaw rate 670 deg/s, hover throttle 0.35), normalized to [-1, 1] as [roll, pitch, yaw_rate, throttle]. Action i is the transition from frame i to frame i+1.
• Episodes have one frame fewer than the source clips so that every frame carries an action.
• The dataset-level fps is the nominal 27; true capture rates vary between 25.6 and 27.7 Hz per clip, and the true per-frame capture times are preserved in the original_timestamp_s feature (seconds since episode start).
• Grayscale frames are replicated to 3 channels for video encoding; the mp4 round-trip deviates from the source frames by about 0.009 mean absolute error in [0, 1] scale.

Features

key	dtype	shape	notes
observation.images.front	video	(3, 128, 128)	grayscale replicated to 3 channels
action	float32	(4,)	[roll, pitch, yaw_rate, throttle], normalized [-1, 1]
original_timestamp_s	float32	(1,)	true capture time, seconds since episode start

Episode task strings distinguish the camera track (forward-facing vs 45-degree tilted) and the camera side (left/right).

License and attribution

The original dataset is released under CC BY-NC-SA 3.0; this derivative carries the same license. Non-commercial use only.

If you use this data, cite the original authors:

@inproceedings{Delmerico19icra,
  author = {Jeffrey Delmerico and Titus Cieslewski and Henri Rebecq and Matthias Faessler and Davide Scaramuzza},
  title = {Are We Ready for Autonomous Drone Racing? The {UZH-FPV} Drone Racing Dataset},
  booktitle = {{IEEE} Int. Conf. Robot. Autom. ({ICRA})},
  year = {2019}
}

See fpv.ifi.uzh.ch for the original data, additional citation requests for specific sequences and ground truth, and full terms.