Skip to content

Robot and Ball Detection

The robot and ball detection is implemented by the module RobotBallDetector. It uses a nano-network based on YoloX1.

Architecture

The input size uses the larger image dimension of the K1 and T1 squared, meaning 544x544 for the K1 and 640x640 for the T1.

Dataset

To be released. We provide two datasets: one with only our own labeled images and a second one which combines the first one with the two datasets of TORSO[^Torso] and whIRLwind4, but cleaned up and only containing labels for robots, balls and penalty marks as well as corrections for those bounding boxes.

Our own dataset was labeled with LabelStudio2.

Training Pipeline

To be released.

  1. https://github.com/Megvii-BaseDetection/YOLOX ↩

  2. https://labelstud.io/ ↩

  3. https://github.com/bit-bots/TORSO_21_dataset ↩

  4. https://intelligentroboticslab.github.io/hslvision/ ↩

Last update: September 11, 2026