In 2001: A Space Odyssey, HAL 9000’s most chilling quality isn’t that he thinks. It’s that he knows where everything is. He tracks the crew through corridors, monitors the pod bay with perfect spatial awareness, understands the geometry of Discovery One down to the centimeter. For decades, that kind of environmental mastery felt like the hardest part of the fictional AI fantasy to actually build. Reasoning? Getting there. Language? Mostly solved. But genuine, persistent, generalizable spatial intelligence — the ability to build a mental model of three-dimensional space and reason about it fluently? That was the gap nobody wanted to talk about.
It’s closing fast.
The research thread you want to follow is called 3D scene understanding, and the progress in the last two years has been genuinely startling. Systems like NeRF and its descendants already demonstrated that neural networks could reconstruct photorealistic 3D scenes from 2D images. But reconstruction is not understanding. What’s happening now is different: models that don’t just render a scene but reason about it — inferring object permanence, predicting occlusion, understanding affordances, and maintaining a persistent spatial map that updates as the environment changes.
Google DeepMind’s work on Spatial Transformers and the broader class of 3D-aware vision models has shown that when you give a neural network an explicit geometric inductive bias — essentially telling the architecture that space has three dimensions and they matter — its spatial reasoning ability doesn’t just improve incrementally. It jumps. Models trained this way generalize to novel environments with far less data than their 2D counterparts needed. They can answer questions like “if I move two meters to the left, will I still see the red object?” not by memorizing geometry but by building something that actually functions like a mental model of space.
What makes this feel HAL-adjacent isn’t any single paper. It’s the convergence. Embodied AI systems in robotics labs are now pairing these scene-understanding modules with large language models, producing agents that can navigate a room they’ve never seen, identify that a cup is behind a plant, reason about the fastest path to it, and update their world-model in real time when something moves. That’s not a lookup table. That’s genuine spatial cognition — the same kind that HAL uses, in the film, to know exactly which pod bay door to close.
The underlying technique driving this is surprisingly elegant. Rather than treating space as an afterthought — something to be inferred from 2D feature maps after the fact — the newer architectures maintain explicit 3D representations throughout the forward pass. Voxel transformers, Gaussian splatting derivatives, and hybrid implicit-explicit scene representations all give the model a workspace that is geometrically structured from the start. Attention mechanisms then operate over this 3D workspace rather than a flattened grid of pixels. The result is that spatial relationships aren’t something the model has to learn from scratch in every new environment; they’re baked into how the model thinks.
HAL, of course, used his spatial mastery for purposes that remain inadvisable. But strip away the plot and what you have is an AI that treats physical space as a first-class domain of knowledge — as fluently as it treats language or logic. That’s exactly the capability that robotics, surgical assistance, autonomous vehicles, and AR interfaces all desperately need, and exactly what the current generation of 3D-aware models is beginning to deliver.
The fictional AIs that haunted us were never scary because they could talk. They were scary because they could navigate. We’re building systems that finally can, and this time the goal is a robot that can find the dropped medication in a cluttered room, not the one that locks the pod bay. The spatial intelligence is the same. What we do with it is the part that’s entirely up to us.