Swarm Robots, World Models, and What Comes Next
Where I place my bets in robotics: world models and 3D maps, self-assembling astrocyte swarms, and the idea that most of the future is already imagined.
When people ask me where to place their bets in robotics for the next ten years, I try to widen the question first. It's tempting to fixate on the robot itself, but some of the most important work is happening around the robot, in embodied AI, in simulation, and in the maps that let a machine understand the world it's standing in.
Look at what NVIDIA is doing with simulation and training environments. Look at a company like Genesis, which just launched its latest humanoid and is doing genuinely amazing things. The robots are getting extraordinary. But here's the catch: you can have the world's best robot, and if it doesn't know there's a table and chairs in front of it and it trips over them, none of that engineering matters.
That's why I'd point people toward something that doesn't sound like robotics at all: the next generation of world models. There's Google Earth, which is satellite imagery. There's Street View, which shows you the street. The next stage is the point cloud of every interior space on the planet, every room, every staircase, every object, because that's the map a humanoid actually needs to operate in our world.
The people building this are worth knowing. Before Google Earth there was a company called Keyhole, which is why Google Earth's geo files are KML, Keyhole Markup Language. Keyhole was founded by John Hanke and Brian McClendon, and it was acquired by Google to create Google Earth. Hanke went on to found Niantic, the company behind Pokemon Go. That gaming side has largely been sold off, and the team is now focused on Niantic Labs, building the next world model. Their technique is Gaussian splatting, one of the engineers there authored the foundational paper on how it works. It lets you take 360-degree images and reconstruct full 3D spaces; it's a bit like LIDAR for finding the point cloud around you, but it works differently, and you can do it from a smartphone just by walking around. I'd argue that this next-generation 3D map of the world, the thing that will let humanoids move through our spaces, is as important as the robotics itself.
Now, if you want to know where I think this all goes, here's my real bet: swarm robotics.
The early glimpse of it is drones flying in formation to draw a shape in the sky, a swarm doing coordinated processing. But the future version is self-assembling robotics. I once worked on a project with the team at NASA Langley that never saw the light of day, built around an idea I called astrocytes.
The name comes from the brain. An astrocyte is a cell that looks a bit like a star, with tentacle-like connections reaching across many neurons at once. There's a wonderful story behind it. When Einstein died, they performed an autopsy on his brain looking for what made him so different, and at first they found nothing. His neurons and synapses looked like everyone else's. His brain was preserved, and roughly twenty years later, someone identified these astrocytes, which act as a kind of superconductor in the brain. They went back to Einstein's brain and found, and don't quote me on the exact figure, something like ten times the number of astrocytes in the region associated with mathematical reasoning and logic. What looked ordinary at first turned out to be extraordinary once we knew what to look for.
My astrocytes robot was a swarm that could build and assemble itself. Each piece would have its own processing and be a fully independent robot, but together they could form larger structures and even build their own tools. I brought the idea to NASA; they were interested in extraterrestrial licensing, using it in space. Picture something that could self-assemble, create the exact tool it needs to fix a problem on the outside of a spacecraft far from Earth, do the repair, and come apart again. I was also looking at more down-to-earth uses, like tooling for large-scale mining and manufacturing.
As AI models get smaller and more lightweight, small enough to run on a single microprocessor rather than a cloud of GPUs, much more of this becomes possible.
I'll end on a humbling thought. I don't actually believe there are many truly new ideas; there's mostly the moment when an old idea finally becomes buildable. The person who invented ultrasound had the core concept back in 1949, there are photographs of a man sitting inside a ring-shaped device doing an ultrasound of his own neck, the first image of a human neck, made with something essentially like a modern scanner. What's changed since isn't the idea. It's AI, miniaturization, and processing power. Most of the future, in other words, is already imagined. The work is bringing it to life.