Reproduce “Robust Navigation for Racing Drones based on Imitation Learning and Modularization” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Tianqi Wang, Dong Eui Chang
A modular navigation system for racing drones, developed by KAIST, combines deep learning perception with model-based control to enhance robustness in dynamic environments. The system, trained purely in simulation, demonstrates superior track completion rates over previous methods and successfully generalizes to unseen photo-realistic textures.
Key insights
- Augmenting visual input with drone state information and explicitly predicting intermediate gate parameters (center, orientation, distance) enhances navigation robustness and explainability.
- Training the imitation learning neural network from scratch on all accumulated data after each DAgger augmentation step significantly improves stability and prevents convergence to local minima.
- Extensive randomization during simulation-based training, particularly for maximum velocity and illumination, is crucial for successful domain transfer to photo-realistic and dynamic real-world environments.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.