Skip to main content

ORB-Net

End-to-end Planning Using Feature-based Imitation Learning for Autonomous Drone Racing

Accepted into ISR europe 2023

Key Contributions​

  • ORB-Net as a neural architecture that could regress motions (translational and yawn velocity) from RGB images and ORB feature maps extracted from a feature extractor (ORB)
  • Technical contributions: Custom Dataloader implemented in PyTorch, used for batching ORB, ORB-mask, RGB images, RGBO to feed into the training pipeline.
  • Imitation Learning pipeline to train the ORB-Net model from an expert PID-controlled agent, via continuous sampling trajectories and mixing policies from both the expert and the agent controller (ORB-Net).

Acknowledgments​

Special thanks to all the co-authors from IVSR Lab and Aarhus University's OpenAirlab. All the simulated trajectories data used in training are from OpenAirlab's Unreal Engine-made Drone Racing Environment.