Stereo vision and planning for the ARTEMIS humanoid robot

For RoboCup 2024 Humanoids league (Winning project)

Features of the proximity package (Inference runs at 60 Hz):

  • Connected Component Analysis for Object Detection: The module performs an infinite component analysis using union-find on the depth map from the ZED stereo camera to detect the k closest objects in the field of view, limited to two objects in this implementation.
  • Feature Extraction & Reference Frame Conversion: It calculates key attributes (width, height, distance, angle) of detected objects and converts these measurements to different robot reference frames as required by other components.
  • Adaptive Thresholding Based on Context: The threshold for detection dynamically adjusts based on the neck angle and field location (e.g., reducing threshold when looking down at corners to avoid detecting the ground or body parts as obstacles).
  • HALO Mode for Vision Failures: In blackout scenarios (e.g., severe occlusion by other robots), it switches to a fallback HALO mode, looking down at the ground at a specific neck angle to identify open navigation paths.
  • Integration with Other Modules: The component is optimized to work seamlessly with the mid-level planner, localization system, and high-level planner to ensure smooth operation in dynamic environments.

The algorithm is as follows:

Algorithm Schematic

[Connected-components with union find algorithm]

Algorithm Schematic

[Infinite Components Analysis Algorithm]

ARTEMIS Image 1

[Close-range object detection on humanoid]