AirHunt

A real-time aerial object-navigation system that connects VLM semantic reasoning with continuous UAV planning.

Overview

AirHunt is a real-time aerial navigation system for large-scale unknown outdoor environments. It accepts natural-language objectives and bridges high-latency vision-language reasoning with low-latency UAV replanning for zero-shot object navigation.

My contribution

  • Designed the high-frequency geometric path planner and the low-frequency VLM semantic-reasoning module.
  • Built the ROS-based communication layer connecting perception, reasoning, planning, and control.
  • Implemented baselines and ablation studies, then built and deployed the custom AirHunt quadrotor platform.

Technical approach

Dual-pathway architecture
Decouples slow VLM inference from fast UAV replanning so that the drone can fly continuously.
Active dual-task reasoning
Selects coverage-aware and task-aware keyframes for efficient semantic reasoning and target verification.
Semantic-geometric planning
Builds a persistent 3D value map and balances semantic priority with geometric travel efficiency.
System deployment
Integrates perception, reasoning, planning, and control through ROS for AirSim and real-world flight.
AirHunt dual-pathway architecture and comparison with sequential and naive asynchronous pipelines
AirHunt system overview. A slow VLM pathway asynchronously updates a persistent 3D value map, while a fast planning pathway continuously turns the evolving semantic memory into flight trajectories. This avoids both the hovering of sequential pipelines and the stale actions of naive asynchronous designs.

Results

73.1%average navigation success rate across 85 simulation episodes
11.6 maverage final distance to the target
120.8 saverage end-to-end flight time
10+ houtdoor real-world flight validation
AirSim navigation sequence from initialization through searching to locating a tent on the beach
AirSim example. For “Fly to the tent on the beach,” AirHunt prioritizes high-value regions, produces a smooth trajectory, avoids revisiting explored areas, and reaches the target.
Three real-world AirHunt navigation tasks in park, landscaped, and residential environments
Real-world examples. The custom quadrotor completes open-set navigation tasks in three challenging outdoor environments, progressing from initialization through semantic search to target localization.

Paper & demos