Familiarity-taxis: A bilateral approach to view-based navigation

Abstract

Many insects use view-based navigation, or snapshot matching, to return to familiar locations, or navigate routes. This relies on egocentric memories being matched to current views of the world. Previous snapshot navigation algorithms have used full panoramic vision for the comparison of memorised images with query images to establish a measure of familiarity, which leads to a recovery of the original heading direction from when the snapshot was taken. Many aspects of insect sensory systems are lateralised with steering being derived from the comparison of left and right signals like a classic Braitenberg vehicle. Here we investigate whether view-based route navigation can be implemented using bilateral visual familiarity comparisons. We found that the difference in familiarity between estimates from left and right fields of view can be used as a steering signal to recover the original heading direction. This finding extends across many different sizes of field of view and visual resolutions. In insects, steering computations are implemented in a brain region called the Lateral Accessory Lobe, within the Central Complex. In a simple simulation we show with a SNN model of the LAL an existence proof of how bilateral visual familiarity could drive a search for a visually defined goal.

Embodied visual route following by an insect-inspired robot

“…This preprocessed image is further cropped into two overlapping sub-images, which are the left and the right views perceived by differential-MB's lateral visual receptive fields. Such low resolution vision is consistent with insect compound eye optics (Schwarz, Narendra, & Zeil, 2011), and proved sufficient to produce good (or even better) performance while reducing computational costs (Steinbeck et al, 2024;Wystrach, Dewar, Philippides, & Graham, 2016).…”

“…Based on the normalised differential MB outputs, the model's steering control rotates its body (a robot) towards the side with a higher familiarity, effectively performing visual servoing. Such a lateral MB architecture, and its potential to produce a suitable steering signal, has been proposed in Wystrach et al (2020), and further studied in Steinbeck et al (2024). However, neither of these works actually used an MB neural network but respectively assumed an ideal visual familiarity function, or used the image similarity of the nearest stored image (i.e., with image location indexed) to assess familiarity.…”

“…Unlike many state-of-the-art visual localisation techniques, the model does not rely on feature detection or deep learning. It essentially performs visual template matching (Labrosse, 2006), for which such low-quality visual inputs are proved sufficient to produce good performance (Steinbeck et al, 2024;Wystrach, 2023;Wystrach et al, 2016;Yihe et al, 2023).…”

“…The idea has been tested in simulation with realistic views in virtual environments based on real ant habitats in two recent works. Steinbeck et al (2024) considers non-straight routes, but uses perfect memory in its model and does not evaluate model performance in closed-loop control. While Wystrach (2023) develops a more biologically realistic model capable of closedloop control, the model is probed in the homing task, i.e., it is expected to return its home in straight lines.…”