Robots Need Ground Truth, Not Just Artificial Intelligence

THE SPATIAL SIGNAL | GROUND TRUTH

October 14, 2026

Spatial AI · LiDAR · Robotics · Digital Twins · Infrastructure

Robots Need Ground Truth, Not Just Artificial Intelligence

“The next breakthrough in robotics may depend less on teaching machines to think—and more on teaching them to understand the physical world.”

The race to develop intelligent robots is exposing an overlooked problem: machines cannot safely work in environments they don't understand. That could give geospatial data an entirely new importance.

On June 25, 2025, a routine sewer inspection in downtown Champaign, Illinois, uncovered something more troubling than a blockage.

A section of iron pipe beneath the city had separated, allowing soil from an earlier excavation to enter the sewer. A void was developing underground, creating the potential for a sinkhole near the intersection of Neil and Clark Streets.

The discovery came from a robotic inspection camera. City employees recognized the danger, brought in a contractor and arranged repairs before the situation deteriorated. The work required the temporary removal of a traffic-signal pole and mast arm.

The camera had performed its job. It showed municipal employees what was happening underground. But people still had to interpret what they were seeing, understand the consequences and decide what to do next. The City of Champaign's account describes a familiar combination of inspection technology and human judgment. Champaign, IL

That gap between observing the physical world and understanding how to act within it is becoming one of the central challenges in robotics.

On Wednesday, October 14, Chinese LiDAR manufacturer Hesai launched a technology intended to help close part of that gap.

Its new Kosmo system combines laser scanning, conventional imagery and 3D reconstruction to capture physical environments and turn them into detailed digital representations. The company believes these datasets will help developers train robots to navigate unfamiliar surroundings, recognize objects and eventually perform more complex physical tasks.

The announcement points to a change in the direction of artificial intelligence.

For years, the industry has concentrated on teaching machines to interpret language, recognize images and generate information. Now it wants machines that can perform useful work in the real world.

For that, intelligence alone isn't enough.

The spatial data problem

Hesai is best known for supplying LiDAR technology to the autonomous vehicle industry. With Kosmo, it is extending that expertise into a potentially much larger market: the spatial data needed to develop and train intelligent machines.

The new system has two components.

Kosmo One is a spatial camera combining high-resolution LiDAR, optical imaging and inertial measurement technology. It captures the geometry, appearance and position of objects within an environment.

Kosmo World is a cloud-based platform that processes captured information into detailed 3D environments. Hesai says these environments can be used for robotics simulation, training and the development of AI world models.

The distinction matters. Hesai isn't simply introducing another scanning device. It wants to provide the infrastructure for collecting, processing and reusing real-world spatial information.

The company has announced partnerships with Chinese robotics developers Galbot and Dexmal. Reuters also reported that a robot developed by Sharpa Robotics had been serving ice cream at a Dairy Queen location in Shanghai, with Hesai crediting its spatial mapping technology for helping achieve a 99% task-success rate.

That figure is company-reported, and serving ice cream in a controlled setting is a very different challenge from maintaining public infrastructure. Nevertheless, it offers a practical illustration of how detailed representations of a working environment can support robotic operations. (Reuters; Hesai) Reuters

The underlying problem is familiar to anyone working in geospatial technology.

A machine must first establish where it is, what surrounds it and how objects relate to one another. Accurate geometry and reliable positioning are essential, but they are not sufficient.

A robot must also understand what those objects are and what will happen if it interacts with them.

That's where the challenge becomes considerably more complicated.

A map can locate a valve. Can a robot operate it?

Consider a municipal pumping station.

The city's GIS identifies the facility and its major assets. Engineering drawings document the pipework. A terrestrial laser scan provides a detailed 3D representation of the interior.

For an engineer or maintenance technician, these records are valuable references.

For a robot expected to perform maintenance, they are only the beginning.

The machine must recognize the correct valve, distinguish it from similar equipment and determine how to approach it. It needs to identify the handle, understand how it operates and avoid nearby obstructions.

More importantly, it must understand the consequences of turning it.

Is the line pressurized? Is the valve operational? Has the equipment been modified since the last survey? Could opening it interrupt service or create a safety hazard?

None of those questions can necessarily be answered by looking at a point cloud.

A conventional 3D model describes the geometry of an environment. A robot performing physical tasks needs additional knowledge about the objects within that environment, their condition and how they behave.

This is one reason researchers are developing world models: predictive representations that help AI systems anticipate how an environment may change in response to particular actions.

An April 2026 research survey, published through arXiv and featured by Microsoft Research, examines how world models are being developed for robotic planning, navigation, simulation and learning.

The research reflects an important distinction. Capturing a realistic 3D environment is not the same as creating a physically accurate simulation of everything a robot might do within it. Microsoft

In a pumping station, the missing knowledge may reside in engineering drawings, maintenance histories, operating procedures, equipment specifications and live sensor readings.

Much of this information already exists within municipal systems. But it is rarely organized with autonomous machines in mind.

The problem isn't necessarily collecting more data. It's making different kinds of information work together.

From inspection to intelligent action

Robotics is already finding practical applications in infrastructure management, although most deployments remain focused on inspection rather than autonomous maintenance.

In September 2025, the Missouri Department of Transportation published research examining robotic bridge inspection technologies.

Researchers evaluated systems across 59 bridges, using drones, structural crawlers, laser scanning, imaging and nondestructive testing. Their objective was to improve inspection efficiency and reduce the need for equipment and operations that can disrupt traffic.

Machine-learning techniques were used to detect, locate and quantify certain defects, including damage to steel bridge components.

The findings were encouraging, but the researchers identified a significant limitation.

The number and variety of defects encountered during the study were insufficient to create the extensive datasets needed for more comprehensive machine-learning-based decision support.

In other words, the robotic systems could collect valuable information, but training machines to interpret the full range of infrastructure conditions remained difficult. (Missouri DOT research, September 2025) DOT ROSA P.

That distinction is important.

A robotic crawler can inspect an inaccessible bridge component. Computer vision can help identify a crack. But deciding whether that crack requires immediate intervention involves structural knowledge, engineering judgment, inspection history and an understanding of risk.

Autonomous repair introduces another level of complexity.

The same progression applies to sewer inspection, utility maintenance and transportation infrastructure.

Robots are becoming increasingly capable of collecting information about physical assets. Their ability to interpret that information reliably—and act upon it without human intervention—is developing more slowly.

For local governments, this is a reason to follow robotics closely, not a reason to assume fully autonomous maintenance is around the corner.

The next customer for municipal spatial data

For local-government GIS and asset-management professionals, the longer-term implications are worth considering.

Cities have invested heavily in mobile LiDAR, street-level imagery, infrastructure inventories and digital twins. These datasets support engineering, capital planning, asset management and infrastructure inspection.

Until now, their principal users have been people.

Robotics introduces another possibility. The same spatial information could eventually help machines understand the environments in which they are expected to work.

Imagine a water treatment facility that has been comprehensively surveyed using terrestrial LiDAR. Engineers use the resulting 3D model to plan equipment replacements and assess proposed modifications.

In the future, portions of that model might also support robotic inspection systems.

Combined with equipment records, operational information and suitable simulation software, it could help developers train robots to navigate the facility, recognize particular assets and reach designated inspection points.

The economic implications could be substantial. A dataset originally commissioned for engineering purposes might acquire additional value as part of the information infrastructure supporting autonomous operations.

But that value would depend on how the information was collected, maintained and made accessible.

A point cloud may reproduce a room with millimeter-level accuracy while revealing little about the equipment inside it. A GIS inventory may identify every pump without providing the detailed geometry required for robotic interaction. A digital twin created two years ago may no longer reflect modifications made during subsequent maintenance.

Even the most sophisticated 3D reconstruction can misrepresent conditions if the underlying information is outdated, incomplete or incorrectly interpreted.

The spatial data governments collect today may become more valuable tomorrow—but only if it can be trusted and reused.

That suggests a practical priority for municipal GIS and asset-management teams.

Accurate asset identifiers, consistent coordinate systems, current condition information, reliable maintenance histories and better connections between GIS and engineering systems may be more important than adding another visually impressive layer to a digital twin.

These are not new concerns. Municipal asset managers have struggled with fragmented records, inconsistent data standards and outdated inventories for decades.

What is changing is the potential value of solving those problems.

Who owns the data that teaches the robots?

There is a commercial question here that local governments should not overlook.

If detailed representations of infrastructure become valuable resources for developing intelligent machines, who controls those datasets?

A municipality might commission a LiDAR survey of a treatment plant, only to discover later that the underlying point cloud is difficult to export from the contractor's platform. Another might purchase a digital twin without securing adequate rights to reuse its imagery, models or associated data.

Such arrangements already complicate asset management and engineering workflows.

They could become considerably more restrictive if the same information acquires value for robotics simulation and training.

Future procurement decisions may therefore need to address questions that extend beyond the immediate mapping project.

Does the agency retain the underlying spatial data? Can it export information in usable formats? Are contractors permitted to use municipal infrastructure models for their own AI development? Does the agreement allow new applications that weren't contemplated when the data was originally collected?

Security presents another complication.

Detailed 3D representations of water treatment facilities, electrical infrastructure and other critical assets cannot automatically be treated as unrestricted training resources.

A model useful for autonomous inspection could also reveal sensitive information about equipment locations, physical access points and facility configurations.

The challenge will be preserving the operational value of these datasets while controlling how they are shared and used.

For public agencies, the immediate lesson isn't to begin collecting information for hypothetical robots.

It is to retain meaningful ownership, access and control over the spatial information they are already paying to create.

The real world doesn't stand still

There is one further obstacle that the robotics industry will have to overcome.

Infrastructure changes.

Pipes corrode. Pavements deteriorate. Bridges develop defects. Equipment is repaired or replaced. Construction crews alter the layout of streets and facilities.

A robot operating in these environments cannot assume that the model used during training still represents what it will encounter.

It needs to recognize the difference between the expected environment and actual conditions.

That requires an ongoing relationship between spatial data, asset records and real-world observations.

For municipal GIS professionals, this introduces an intriguing possibility: maintaining accurate representations of infrastructure could eventually become part of keeping autonomous systems safe and operational.

A digital twin would no longer be valuable simply because it provides a comprehensive picture of an asset at a particular moment. Its usefulness would depend on how accurately it reflects changes over time—and how effectively that information can be communicated to machines.

This is a demanding technical challenge. It also raises questions about data validation, update frequency, liability and the level of human oversight required when robots work around essential public services.

Hesai's Kosmo launch does not resolve those questions. Nor does it demonstrate that municipal infrastructure can now be maintained autonomously. The company is introducing technology intended to help developers collect and process the spatial information needed to advance robotics.

The significance lies in the direction of the technology.

As intelligent machines move from controlled demonstrations into more demanding physical environments, the quality of their information about those environments becomes critical.

For local governments, that doesn't mean rebuilding GIS systems around speculative robotics applications. It does suggest looking differently at the spatial information they already own.

Accurate asset inventories, well-maintained 3D models and reliable connections between location, equipment and maintenance records could become increasingly valuable as machines acquire the ability to perform physical work.

The geospatial profession has spent decades developing ways to measure, document and understand the physical world. Much of that expertise may prove essential to the next stage of artificial intelligence.

We have spent years teaching machines to think. Now we must teach them where they are, what surrounds them and what happens when they act.

That is a very different challenge.

And it starts with ground truth.

· · ·

Ground Truth is The Spatial Signal's editorial perspective on the technologies, decisions and developments changing how we understand and manage the physical world.

Reporting & Research

Reuters | October 14, 2026 China's Hesai launches spatial camera to boost AI-powered robotics adoption

Hesai | October 14, 2026 Building the Data Infrastructure for Physical AI: Hesai Launches Kosmo 3D World Data Engine

City of Champaign | July 31, 2025 Sanitary Sewer Televising Prevents Future Infrastructure Issues

Missouri Department of Transportation | September 2025 Traffic Disruption-Free Bridge Inspection Initiative with Robotic Systems

Research | April 2026 World Model for Robot Learning: A Comprehensive Survey