We developed a system to attach ants on top of an air floating ball, acting like a treadmill. The ants are remarkably at ease on the setup, whether directly in the field or in the lab virtual reality system: they maintain their motivation to navigate and display their usual motor behaviours, as if undisturbed. Field experiments using these trackball systems enabled us to ask novel questions to dig into their mechanisms. For instance, we reveal that, contrary to previous theories, ants could recognise their route no matter the direction they faced, and could infer from this situation which side was the correct goal direction. Following on this, we showed that learning is not something that happen during a specific moment, such as when encountering a reward or looking at the goal, but happens continuously, echoing the so-called ‘latent learning’ in vertebrates. Also, we revealed that navigating ants are using a two-step strategies: visual recognition of the scene is not used to drive motor behaviour directly, but this information updates a neural representation of the goal direction, based on celestial cues, which they then use to guide their movements.
Our modelling effort revealed how the insects brain circuitry could implement these mechanisms. We revealed how simple neural process in the insect early visual system could enable to strongly improve the recognition of complex natural scenes, which happen deeper in the brain. This compression of the visual information explains how ants can learn continuously while navigating over thousands of square meters, without memory saturation. Further experiments in the field revealed how ants learn aversive memories to avoid regions associated with danger, and our neural models show how such aversive memories can be combined with appetitive memories during navigation. When embodied in a simulated agent navigating in reconstructed worlds, our neural models now achieves amazingly robust navigation!
Regarding the motor aspect of the ant’s navigation skills, we revealed the existence of an intrinsic oscillator at the core of their navigation system, and located in an ancestral pre-motor area of their brain. We show that multiple pathways from different modalities converge onto this oscillator, which provides a natural way of integrating them while maintaining a smooth and efficient behavioural output.
In parallel, we developed a Virtual Reality system based on LED’s, to project reconstructed natural worlds on a cylindrical screen, in the centre of which the ant is navigating on its treadmill. This setup enable to manipulate the scene while the ant is navigating. For instance, by reversing the relationship between the ant’s movement and the visual feedback, we showed that ants compute predictions of the motion of the scene they expect based on their own movements. They do not adjust their locomotion using the perceived motion, but using a prediction error, that is, the mismatch between their actual perception and prediction, like vertebrates’ brain do. What’s more, the way ants make prediction is not trivial: they combine information from their motor commands, their proprioception as well as the appearance of the scene (some scene ‘should’ produce more optic-flow when turning than others, and ants take this into account).
Here again, neural modelling enabled us to explore how these mechanisms can be implemented in the insects’ brain. Notably, we showed how neural feedbacks from motor commands, intrinsic oscillator, and learning centres can interact to orchestrate the continuous formation and retrieval of the multiple parallel memories used to produce a robust navigation behaviour.
