SDFGen: A library to generate Signed Distance Fields

Robotics intern @ Mainblades (3 months)

Overview

Mainblades does aircraft inspections autonomously with drones. During my internship, I built a library (based on Voxfield) to represent the obstacle map with SDFs (Signed Distance Fields), this enables using gradient based path planning algorithms, generating smoother trajectories.

Point cloud reconstruction SDF from point cloud
Point cloud reconstruction SDF from point cloud

Contributions

  • Built SDFGen, to convert LiDAR point clouds and reconstructed aircraft models into SDFs (signed distance fields) maps.

  • Developed sdfgen_live (ROS) to integrate live LiDAR/rosbag data during flight and publish an SDF map for online planners while supporting initialization from prior maps for faster startup.

  • Validated with static reconstructed aircraft models (B747/B777), flight rosbags, and hardware-in-the-loop on a DJI M300 and achieved 2 Hz SDF generation onboard.

Languages and tools used: C++, ROS