Overview
- Bee-Nav, detailed Wednesday in Nature, runs neural networks between about 3.4 and 42.3 kilobytes.
- The drone first makes a short learning flight to store panoramic views that its tiny model maps to a homeward direction and distance.
- The approach uses path integration to reach the learned area near the start, then the visual memory guides the final return.
- Tests scaled from large indoor hangars to a 600-meter outdoor flight on a Raspberry Pi 4, with indoor trials succeeding in every run.
- Wind made the drone tilt and warped its images outdoors, which cut success to about 70 percent and flagged needs such as multi-site memory, landmark-poor starts, and local obstacle avoidance.