science
Bee-Nav: Tiny Neural Network Inspired by Honeybee Learning Flights Enables Visual Homing in Simulation and Limited Real-World Drone Tests
A Nature research paper proposes Bee-Nav, which trains a small neural network during robotic learning flights to map omnidirectional images to a home vector. Real-world drone tests returned to within 0.5 m of home for all 30–110-m flights using a 3.4-kB network. The approach requires training on just 0.25–10.00% of the total flight area in simulations for realistic path integration noise.
May 12, 8:00 PM·2m read2 sources
NA
Hoover InstitutionAudio version
Tap play to generate a narrated version.
Developing·Limited corroboration so far. This page will refresh as more sources emerge.
A Nature research paper proposes a navigation strategy called Bee-Nav inspired by honeybee learning flights. Bee-Nav uses a tiny neural network trained during robotic learning flights to map omnidirectional images to a home vector based on path integration. Honeybees perform one to several short learning flights before departing on longer exploration or foraging trips.
Honeybees can fly several kilometres from their hive. The learning flight trajectory fits in a 10-m-radius circle around the home location. 00% of the total flight area.
In simulation 1,000 foraging flights were tested using a block-wave outbound search pattern followed by straight inbound trajectories based on noisy path integration.
6%. 4%.
Path integration drift in the real robot was measured using a downward-looking optical flow sensor, laser for height, and gyro for heading. 5 m of home for 100% of 30–110-m flights.
In real-world experiments a small drone achieved 70% success for 200–600-m flights in windy conditions. The small drone used a 42-kB neural network for 200–600-m flights. 3-kB attention neural network was used in NVIDIA Isaac simulator visual homing experiments.
In ten different simulated forest environments Bee-Nav achieved 100% success rate within the LHA radius. 5 times the LHA radius depending on the simulated environment. Within the LHA the neural network angular errors were smaller than about 40° and distance errors smaller than 2 m.
