Reproduce “AirDreamer: Generalist Drone Navigation with World Models” 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
Zian Liu, Andong Yang, Chunkai Yang, Ruidong An, Chao Gao, Guyue Zhou
AirDreamer helps a drone with a forward-facing depth camera navigate hidden obstacles by using a recurrent learned scene representation to choose velocity and yaw, and by training with imagined rollouts. In XTDrone, it achieved a reported 59.3 ± 15.8% success rate versus DepthNav’s 54.0 ± 13.9% across 150 runs; real-world evidence remains qualitative.
Key insights
- Decoupling environment modeling from policy learning enhances robustness to out-of-distribution scenarios and improves generalization in cluttered navigation tasks.
- Sparse reward functions, combined with a world model capable of generating long-horizon imagined rollouts, enable the emergence of complex, adaptive behaviors such as active visual scanning and local optima escape.
- Domain randomization of drone dynamics is critical for effective sim-to-real transfer, allowing policies trained purely in simulation to perform reliably on physical hardware.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.