Building the Spatial Infrastructure for General-Purpose Robotics

Niantic Spatial’s CEO, John Hanke, recently wrote about the opportunity for AI beyond the screen: not just in generating content or automating workflows, but in transforming the much larger portion of the economy that exists in the physical world.

The rise of robots designed to reason and move around the world more like humans than machines signals that this shift is underway.

Robots have been part of the physical economy for decades. Industrial arms weld and assemble car parts, mobile robots move pallets across warehouses, and inspection drones scan plants for corrosion. Large parts of global industry already run on this category of robotics.

What’s changing is not the presence of robots, but their scope – what they’re capable of and where they can go.

For most of their history, robots have been tightly specified systems designed for narrow environments. Their productivity has been limited by their programming. The ambition emerging now – fueled by the convergence of increasingly articulated, higher-DoF hardware, and foundation models – is to loosen that coupling.

The goal is not just to automate one-off tasks, but to build robots that can navigate and operate in the same complex spaces, and in the same complex ways, that humans do.

That ambition collides quickly with one of robotics’ unsolved problems: navigation in human environments.

Urban environments are an unforgiving operating domain

Cities are ever-changing obstacle courses. GPS signals degrade unpredictably between buildings, environmental conditions shift gradually and then abruptly, and small inconsistencies – a blocked curb cut, a partially blocked pickup zone – materialize constantly. Humans navigate these without thinking. Robots can’t.

Coco Robotics, which operates one of the largest urban fleets in the world, is leading the industry on tackling these challenges. It has made substantial progress in autonomy and safety, informed by its wealth of real deployment experience. This is why Coco is an ideal partner for us as we seek to refine our infrastructure for the rising generation of robots in human spaces.

Our role in the ecosystem is to strengthen one specific layer that these robots need: geospatial intelligence. This begins with precise, resilient visual localization.

If robots are to move intelligently among people, they must know where they are in the world and where they’re headed. They must localize with enough consistency and precision to support behavior that feels intuitive and human, rather than mechanical and inanimate. This matters not just for operational success, but for social success – a robot working in public is scrutinized not just on whether it completes its task, but on how it moves.

That requirement is not unique to delivery.

Spatial challenges are converging across embodiments and environments

The constraints that surface in urban deployments reappear in other areas. Humanoids navigating to the right workstation in a large industrial facility must maintain a stable global pose to get there, and to be oriented correctly before starting a fine-grained task. Quadrupeds inspecting oil refineries or construction sites must be resilient to degraded signals and trusted to move safely among humans.

The hardware varies, but the geometry of the world does not. Without reliable grounding in space, even the most capable embodied policy model breaks down.

Our localization service, VPS, provides that grounding live. Our reconstruction service, which produces 3D Gaussian splats and meshes of real-world locations, enables the same thing in simulation.

Today, robotics companies have to make do with training against vast quantities of synthetic data. These approximations of real-world environments create approximations of desirable real-world behaviour, which break down once you move from screen to street.

The sim-to-real gap remains a constraint on scaling autonomy. Our reconstruction service offers a way to narrow it by grounding learning in metric-scale, high-fidelity 3D visualizations of the places robots will be or are already deployed – not imagined versions.

Reconstruction also changes the human-machine dynamic. Even as autonomy advances, human oversight will remain part of deployments for many years to come. A crisp, three-dimensional model of operating environments provides materially better context for operators and engineers.

This is particularly potent with the addition of open, three-dimensional semantic understanding. Adding the ability for robots to know the world before they enter it – with a level of detail and intuitive querying that has never been possible before – unlocks true, shared geospatial intelligence across a fleet.

And it’s only possible with a geospecific – not geotypical – model of the world.

Real-world robots need a real-world map

Throughout tech history, as vertical markets have matured, certain parts of the stack have naturally emerged as making more sense to outsource than to build from scratch.

The layers that ultimately become horizontals – think payments, cloud computing, and mapping – are expensive to build and maintain, improve with usage across companies, and aren’t the core product the market provides to customers.

As deployments of robots in human spaces are scaling, the underlying geospatial challenges are beginning to converge. And it’s going to take more than general-purpose policies and a two-dimensional understanding of space for artificial intelligence to overcome those.

It will require infrastructure that lets embodied AI reason and act within the constraints of the real world – not an approximation of it. That’s what we’re building at Niantic Spatial with our Large Geospatial Model: a living model of the world that humans and machines can talk to.

General-purpose robots are coming. Their success depends on a map as real as the world they inhabit.