How Do Bees See the World? A Normative Visual Reinforcement Learning Model for Insect Navigation.

• Applicant: Professor Dr. Andrew D. Straw
• Subject Area: Cognitive, Systems and Behavioural Neurobiology
• Term: since 2026
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 569980362

Bees and other central place foraging insects are among the master navigators of the insect world. To survive, they must rely on their sensory apparatus and a tiny brain to memorize food locations and successfully navigate between those and their nest. Solving this complex task on very limited computational hardware suggests efficient and highly specialized neural processing of the sensory input. In this project, we focus on vision as the main sensory modality and use a normative modelling approach to investigate which internal neural representations of the raw visual input are – from a theoretical and modelling perspective – most effective for performing these critical navigational tasks reliably and under a range of visual and ecological conditions. We simulate insect navigation using an insect-inspired deep Reinforcement Learning (RL) model trained on real-time realistic ‘bee-eye’ renderings of a virtual reconstruction of actual experimental field sites. Layers of the model roughly correspond to the visual processing pathway from the retina though the optic lobes to the mushroom bodies, the main learning-associated neuropils in the insect brain where RL type learning is likely to be implemented biologically. In this project we employ deep RL as a normative model for neural processing in the insect brain and focus on the following questions: Which representations of the visual input will the agent learn in each successive layer of the model, when trained on a naturalistic navigation task in an RL paradigm without additional supervision? Which early visual feature extraction will the agent learn to enable robust navigation under changing conditions? How do ecological and optical conditions during training, for example of diurnal vs nocturnal species, shape the visual features the agent learns? The insights gained from such a normative approach will be relevant for neuroscientists studying visual processing and/or navigation, providing experimental guidance for neuroanatomical, electrophysiological, and behavioral studies, as well as computer scientists and roboticists interested in insect-inspired visual processing supporting efficient and robust navigation.

• DFG Programme: Research Grants