Think Like a Pilot: Fine-Grained Long-Horizon UAV Navigation
Researchers have introduced FLIGHT, a new benchmark designed for fine-grained, long-horizon Unmanned Aerial Vehicle (UAV) navigation tasks. This benchmark aims to bridge the gap in existing Vision-Language Navigation (VLN) and Vision-Language-Action (VLA) tasks by incorporating multi-stage instructions and detailed 6-DoF trajectory annotations. To enable real-time reasoning and precise control, they also proposed FLIGHT VLA, an asynchronous architecture that separates a VLM for task reasoning from a diffusion model for continuous action control, guided by explicit "Pilot Reasoning" texts. AI
IMPACT This benchmark could accelerate research into more sophisticated autonomous drone behaviors and real-world applications.