Overview
- Bee-Nav, detailed Wednesday in Nature, demonstrated homing from as far as 600 meters in outdoor tests.
- The drone first learns a panoramic neighborhood, then uses path integration for the long return and switches to learned visual cues near home.
- A short learning flight records surrounding views that a tiny onboard neural network maps to home-pointing vectors to steer the final approach.
- The system ran on a Raspberry Pi 4 with neural networks as small as 3.4 to 42.3 kilobytes, replacing bulky map building.
- Indoor hangar trials reached perfect success, while windy outdoor runs fell to about 70% because tilt warped the panoramic images the model reads, with researchers eyeing greenhouse monitoring and light inspections once robustness improves.