LIDo banner

Recent projects

Apply now

Find out more about the different routes to entry and our eligibility criteria

Find out more

Marco Abrate: From movement to cognitive maps: AI models show how learning to move shapes hippocampal spatial coding

rats
From movement to cognitive maps: AI models show how learning to move shapes hippocampal spatial coding

From movement to cognitive maps: AI models show how learning to move shapes hippocampal spatial coding.

A new study – led by Marco Abrate, Caswell Barry, and Thomas Wills at UCL, and published at the International Conference on Learning Representations (ICLR 2026) – reports that changes in early locomotor statistics play a key role in the emergence of spatial representations in the hippocampus, a brain region essential for memory and navigation.

Using published tracking data from rats aged postnatal days (P) 11-25 exploring open-field arenas, the team characterised each animal’s movement through distributions of speed and rotational speed, along with transition probabilities between spatial bins. Rather than assuming that spatial understanding emerges solely from hardwired genetic programmes, the researchers asked whether the gradual maturation of locomotion – from crawling to coordinated running – might itself drive the development of hippocampal spatial representations.

By computing Jensen-Shannon distances between individual locomotor profiles and applying Gaussian mixture modelling, the team identified three distinct developmental clusters. In "crawl" (median age P13.5), pups moved with slow, poorly directed locomotion. By P16 they progressed to "walk," with more stable and directional movement. By P20, "run" emerged – fast and coordinated, resembling adult locomotion. The team then trained shallow recurrent neural networks (RNNs) – models that maintain an internal memory over time through recurrent connections – sequentially on simulated trajectories matching each stage, with each network inheriting the weights of the previous stage to mimic continuous development. At each timestep, the network received an egocentric (self-centred) panoramic visual frame alongside vestibular signals encoding speed and rotational speed, with the objective of predicting the next visual frame. No spatial information was provided – any internal map had to emerge from solving the task alone.

What emerged was striking. Hidden units that respond to specific locations appeared in the network in the same order as in the brain. Cells encoding head direction appeared first, then place cells encoding location and finally, when adult-stage grid cell inputs were added, cells resembled a fully formed brain’s navigation system. This sequential emergence mirrors the known biological timeline of how these neurons develop in young rats.

The model also generated a testable prediction: that conjunctive place-direction tuning – encoding both location and the animal's heading simultaneously – increases during the same developmental timeline, which the team confirmed by analysing CA1 hippocampal recordings. A key validation showing the model anticipates new biological observations.

The work provides a mechanistic link between early physical experience and hippocampal spatial neuron development. Critically, control experiments showed that simply increasing the temporal gap between visual frames – mimicking faster sensory sampling without changing the full movement statistics – was insufficient to reproduce this development. The specific locomotor statistics of each developmental stage, not merely increased sensory change, were necessary to drive the emergence of allocentric (world-centred) spatial representations.

Beyond basic neuroscience, the findings carry a broader message: how you move through and experience the world shapes how your brain develops. This principle may extend well beyond rats – hinting that the richness of early physical experience plays a fundamental role in wiring the brain for spatial awareness and navigation in all animals, including humans. Understanding this process could shed light on developmental conditions affecting spatial awareness, and may also inform the design of AI systems that learn to navigate through interaction with their environment, rather than relying on pre-programmed representations.