Robots Grew Grid Cells. Your Brain Had Them First.

Spend ten minutes watching a two-year-old explore a new playground. They start near the entrance, push toward the swings, loop back past the slide, pause to investigate a patch of gravel, then find the swings again — and this time they head there directly, without retracing their steps. They just built a map.

No GPS. No training data. Just a few minutes of stumbling around, and suddenly they know where things are relative to each other in space.

I find this miraculous. I also find it deeply useful to think about when I'm building robots — because this exact problem, constructing an internal map of a space from movement alone, is one of the hardest things we ask machines to do.

Here's the punchline: the brain has a dedicated neural system for this. It runs on hexagonal geometry. And when a team at DeepMind trained a deep RL agent to navigate a virtual maze — no biological constraints, no prior knowledge, just trial and error — it independently grew the same hexagonal structure. Evolution and machine learning converged on the same solution.

That's not inspiration. That's convergent engineering.

The GPS Nobody Designed

In the 1970s, John O'Keefe discovered "place cells" — neurons in the hippocampus that fire whenever an animal occupies a specific location. Each cell claims a territory. Together they tile the environment like a mosaic. You could, in principle, read out exactly where an animal is just by watching which place cells are firing.

Then in 2005, Moser and colleagues discovered something stranger: grid cells in the medial entorhinal cortex. These neurons fire in a hexagonal lattice pattern — evenly-spaced fields arranged in perfect triangles across the entire navigable space. As you move, different grid cells activate, encoding position with a modular, overlapping precision that's extraordinarily efficient.

According to Dong and Fiete (2024), grid cells are now understood as a near-optimal solution to path integration — tracking your position from movement signals (speed, direction) without needing landmarks. The hexagonal pattern maximizes spatial coverage while minimizing the neuron count. It's elegant engineering that evolution stumbled into, not a blueprint anyone drew up.

Children develop this system gradually. Infant navigation starts egocentric — "the toy is to my left" — and transitions toward allocentric, map-like representations as the hippocampus matures through early childhood. Around age 2–3, genuine cognitive mapping appears: the ability to take shortcuts, infer the location of unseen objects, hold a spatial framework in mind while you're not actively moving. That direct line to the swings? That's allocentric navigation clicking into place.

The Robot That Grew Grid Cells

Here's where it gets genuinely exciting. The 2018 DeepMind navigation study — reviewed by Dong and Fiete (2024) — trained a deep RL agent to navigate virtual environments with one incentive: find the goal. No hexagonal structure in the loss function. No geometry built in. Just rewards for successful navigation.

The agent independently developed grid-cell-like representations in its internal layers.

This is not the brain "inspiring" AI in a vague, metaphorical sense. This is the same solution — the same geometry — emerging from the same pressure. Two completely different systems, biological and artificial, asked to solve the same navigation problem, landing on hexagonal lattice codes.

What does this tell us? The grid code probably isn't a quirk of vertebrate biology. It looks more like a mathematical near-necessity — a structure that any sufficiently pressured learning system will discover when the task demands efficient path integration. Evolution found it over millions of years. Deep RL found it in training runs.

One Map for Everything

Haga et al. (2025) push this even further. In a PNAS paper, they propose a unified computational model — Disentangled Successor Information (DSI) — that explains both place cells and grid cells in spatial navigation and concept cells (neurons responding to high-level semantic categories) within a single mathematical framework.

The provocative claim: there's a formal mathematical equivalence between the value function for goal-directed spatial navigation and the information measures used in word-embedding models. The brain might be using the same geometry to navigate a room that it uses to navigate an idea space.

This isn't metaphor. Concept cells — neurons firing in response to "the Eiffel Tower" or "grandmother" — have been found in the hippocampus, the same region as place cells. The DSI model produces grid-like population codes in navigation tasks and word-embedding-style geometry in semantic tasks. Same computational substrate, same code, different domain (Haga et al., 2025).

For robotics and AI practitioners, this is a fascinating design prompt: maybe systems that handle physical navigation and abstract reasoning will eventually converge on the same internal representational geometry. Not because we planned it. Because it's the right structure for both problems.

The Offline Rehearsal Problem

Navigation isn't just about knowing where you are. It's about planning where to go before you move.

Papale and Buffalo (2025) review the accumulating evidence for awake replay in the hippocampus — place cell sequences reactivating during quiet rest, not just during sleep. Critically, these replays are forward-directed during goal-seeking: before an animal moves toward a goal, its hippocampus is already running the path. The simulation happens before the body commits.

Children do this behaviorally, even if we can't watch their hippocampi directly. A four-year-old pausing at the corner before heading to the park isn't just hesitating — they're computing. That small delay is prospective simulation.

In robotics, we call the equivalent look-ahead simulation or model predictive control: plan a trajectory by mentally playing out consequences before executing. Most robot navigation systems implement this with explicit planning algorithms. The hippocampus does it with spontaneous neural reactivation. Same function, radically different implementation.

And there's one more piece: even before the plan, the body needs to predict the physics of the scene. Pramod et al. (2025) showed that parietal and frontal regions of the human brain encode the contact relationships between objects — support, containment, attachment — and generate forward simulations of how physical scenes will evolve. The brain maintains a dedicated physics engine, and it runs continuously as we navigate space.

Children developing navigation ability aren't just learning where things are. They're building an integrated system: grid codes for position, replay for prospective planning, physics simulation for predicting what happens when they touch things. That's the full stack.

What This Means If You're Building Things

A few things fall out of all this for anyone working in spatial AI or embodied robotics:

Grid structure can emerge — let it. Deep RL agents can grow grid-cell-like representations under navigation pressure. If your agent isn't developing structured internal geometry, it might be a signal that your training environment doesn't demand genuine path integration.

Space and concept may share an architecture. If Haga et al. (2025) are right, a unified representational scheme handles physical navigation and abstract semantic relationships. That's a hint toward building general-purpose agents that don't need separate systems for moving through rooms and moving through ideas.

Offline simulation is load-bearing. Children and animals extend spatial knowledge through replay and forward simulation — not just active movement. Agents that only learn from trial-and-error during active exploration are leaving resources on the table. Awake replay in the hippocampus is computationally active, not passive noise (Papale & Buffalo, 2025).

Embodiment isn't optional. The two-year-old at the playground built their map by moving — integrating proprioception, vestibular signals, visual flow. Pramod et al. (2025) showed the brain's physics engine runs on sensorimotor predictions about contact and consequence. That's not a vision problem or a language problem. It's a body problem. Text-only systems don't have this path available to them, which is why spatial reasoning remains one of the clearest gaps between embodied biological cognition and disembodied AI.

The two-year-old solved it in ten minutes. We're still working on it. But the fact that two completely independent systems — mammalian evolution and deep reinforcement learning — converged on the same hexagonal geometry means we're asking the right question. The answer exists. We've found part of it. Now we just have to build the rest.