Reproduce “FlowPilot: Real-Time World-Action Modeling for Agile UAV Navigation” 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
Runqing Wang, Ding Yu, Pengyuan Min, Xinhong Zhang, Wei Xiao, Yu Hu, Jie Chen, Fu Zhang and 1 more
FlowPilot, developed by a collaboration including the Beijing Institute of Technology, introduces a compact world-action model for agile UAV navigation, achieving sub-18 ms latency on embedded hardware using a dual-stream network. It generates smooth, C2-continuous trajectories via Bernstein polynomials and demonstrates robust zero-shot transfer in real-world cluttered environments at speeds up to 5.5 m/s, outperforming baseline methods in simulation.
Key insights
- Synchronously coupling future-depth prediction with action denoising within the dual-stream model significantly enhances navigation safety and traversal efficiency by providing anticipatory guidance.
- Utilizing Bernstein polynomials for action representation provides a compact, inherently smooth, and state-consistent trajectory that is directly executable by agile flight controllers.
- A two-phase training approach, starting with world-modeling pretraining before full world-action alignment, effectively trains a robust policy with broad environmental understanding and precise action generation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.