Direct learning of home vector direction for insect-inspired robot navigation

TL;DR

The paper tackles the problem of homeward navigation by insects by directly learning, from omnidirectional visual percepts acquired during learning flights, a 2D unit-vector that points toward the nest relative to the current gaze. It achieves this with a compact CNN that outputs the $x$ and $y$ components of the unit vector $(u_x,u_y)$, trained on rectified views across 360 gazes per image and evaluated in both simulated forests and real-world environments. Key findings show average directional errors well below the 90° threshold for successful homing and often below $24^{\circ}$ under textured illumination, with the learning trajectory near the nest critically influencing performance; the approach also enables odometry-drift correction during long-range homeward travel and demonstrates viable homing control of a simulated quadrotor. The work suggests a memory-efficient, scalable visual-homing paradigm with potential applications in large-scale robotic navigation, particularly where odometry drifts are a concern and explicit map-building is undesirable. Future work includes real-robot deployment, online learning, distance estimation, and comparisons with SLAM-based relocalization methods, aiming to further close the gap between biological inspiration and robust autonomous navigation.

Abstract

Insects have long been recognized for their ability to navigate and return home using visual cues from their nest's environment. However, the precise mechanism underlying this remarkable homing skill remains a subject of ongoing investigation. Drawing inspiration from the learning flights of honey bees and wasps, we propose a robot navigation method that directly learns the home vector direction from visual percepts during a learning flight in the vicinity of the nest. After learning, the robot will travel away from the nest, come back by means of odometry, and eliminate the resultant drift by inferring the home vector orientation from the currently experienced view. Using a compact convolutional neural network, we demonstrate successful learning in both simulated and real forest environments, as well as successful homing control of a simulated quadrotor. The average errors of the inferred home vectors in general stay well below the 90° required for successful homing, and below 24° if all images contain sufficient texture and illumination. Moreover, we show that the trajectory followed during the initial learning flight has a pronounced impact on the network's performance. A higher density of sample points in proximity to the nest results in a more consistent return. Code and data are available at https://mavlab.tudelft.nl/learning_to_home .

Figures (7)

• Figure 1: During a learning trajectory (1), we capture omnidirectional images (2) and train a convolutional neural network (3) to estimate the home vector direction in a simulated or real environment (4).
• Figure 2: CNN with kernel and feature map sizes and example input. Convolutional layers are shown in teal; fully connected (FC) layers in blue. The gaze is indicated as a black or white arrow and the home vector as a red arrow. The last nodes have an output value in the range of $[-1,1]$, encoding the coordinates of the relative home vector. Note that the presented gaze angle $0^\circ$ corresponds to north/upwards.
• Figure 3: Simulation environment and training performance for (A) a $10\times10$ m grid and (B) an Archimedean spiral. The bearing maps show the predicted home vectors for all sampled locations, with the nest location in red and trees shaded in green.
• Figure 4: Generalization of networks trained on (A) a $10\times10$ m grid and (B) an Archimedean spiral to unseen locations, showing predicted home vectors and resulting stream plots. Locations seen during training are in magenta, the nest is in red and trees are shaded in green.
• Figure 5: Analysis (following selvaraju2017gradcam) of the highest gradients of the network output with respect to the activations of the second convolutional layer when trained on (A) a $10\times10$ m grid and (B) an Archimedean spiral. Predicted home vectors after training (A.1/B.1) change when evaluated in an environment with a landmark array rotated $90^\circ$ CCW (A.2/B.2). Original landmarks are in green, rotated ones in teal, and dominant landmarks are marked with an asterisk.