Over the past year, spatial intelligence has evolved from an academic concept into one of the hottest keywords in the AI circle. At the end of 2025, Fei-Fei Li published an article calling it "the next frontier of AI," comparing language models to "word craftsmen in the dark": eloquent but lacking experience, knowledgeable but unable to stand on solid ground. In 2026, World Labs completed a $1 billion funding round, and the chief scientist of Qunhe Technology predicted that "the explosion of spatial intelligence will take only 24 months."
Capital is pouring in, papers are being published, and concepts are being repeatedly cited. But if you ask, "What does spatial intelligence actually look like?" most people's answers still remain at the model layer: an AI that can generate 3D scenes, an algorithm that can understand spatial relationships, or an agent that can navigate in virtual environments.
These answers are not wrong, and they are also concepts frequently highlighted in news headlines, but they overlook a more fundamental question: What can spatial intelligence actually do? How is it closely related to our production and life?
On October 14, Hesai Technology officially released the 3D world data engine Kosmo and simultaneously launched the world's first 8K ultra-high-definition AI spatial camera, Kosmo One.
(Xiang Shaoqing receives and officially releases Kosmo One from the hands of the Sharpa D01 robot)
From the Spring Festival Gala robots to the animated films screened by Pixar and the phenomenal Black Myth: Wukong, Kosmo has broken through the boundaries between fields, enabling the real world to be "copied" into assets that AI can understand and train on.
Physical AI, embodied AI, spatial intelligence... Over the past year, these three concepts have taken turns appearing in the headlines of the AI industry, with each wave of hype higher than the last. But despite the buzz, few people have clearly explained the relationship between them.
What physical AI aims to do is to bring AI out of the screen and into the real world, to drive a machine, manipulate an object, and complete a handling task. Embodied AI is the most typical form of physical AI: giving AI a body so that it can perceive, make decisions, and act in real space. Spatial intelligence, on the other hand, is the prerequisite for this body to function properly.
It can be said that physical AI is the goal, embodied AI is the carrier, and spatial intelligence is the foundation.
The progress of embodied AI over the past two years is obvious to all. Humanoid robots can run and jump, dexterous hands can pick up eggs, and end-to-end models have completed a series of impressive operation demonstrations in the laboratory. At the 2026 Spring Festival Gala, robots performed martial arts and acted as backup dancers, stunning the audience with their smooth performances. The release of GPT-6 added fuel to the fire: in some tests, Astra directly output the pose of the robot's end effector, controlling the robotic arm to put blocks into a bowl, succeeding 19 times out of 20, with a success rate even exceeding some policy models specifically trained for robots. General large models are crossing the once clear boundary, reaching out from the digital world to the physical world.
However, performances cannot replace actual work in real physical life. Without spatial understanding capabilities such as position, orientation, scale, and mutual relationships, robots cannot even judge that "the cup is 30 centimeters to the left of the table," let alone pick it up.
To train spatial intelligence, the first requirement is real-world spatial data, that is, using data to build a map. This type of data is precisely the most scarce resource at present. 2D images, simulation data, and real-machine teleoperation cannot meet the needs of the rapid development of embodied AI.
The direct consequence of this bottleneck is: robot hardware is advancing, models are advancing, but the data infrastructure connecting the two is almost blank. GPT-6 can do a good job with policy models, but it cannot get enough high-quality real spatial data for training. Fei-Fei Li has drawn the blueprint for spatial intelligence, and World Labs has created models for generating 3D scenes. But for these capabilities to be implemented on a real robot, what is missing in the middle is a complete set of underlying pipelines from collection to processing to output.
Ultimately, what AI lacks is a world map marked with physical spatial positional relationships. Without this map, embodied AI can only spin in a blank void.
The lack of a map has stalled spatial intelligence, and spatial intelligence has stalled embodied AI. When this chain is transmitted to the very end, it results in robots that look smart in demonstration videos but become confused in real-world scenarios.
It is not that no one in the industry has tried to draw this map. But most players only occupy one end: those doing simulation ignore realism, drawing imaginary maps with mountains, rivers, streets, and buildings all present, but severely distorted; those doing real-machine collection ignore scale, drawing rubbing maps, rubbing inch by inch from the real world, which is too slow, too expensive, and too small; those doing 3D reconstruction ignore semantics and trainability, having only contour lines; those doing joints do not touch data, having no map at all.
The result is that embodied AI companies have to switch back and forth between multiple tools, piecing together a barely usable route. There is not a single map that can take them from start to finish.
GPT-6 has shown the industry the possibility of general models "taking action," and Fei-Fei Li has turned spatial intelligence from an academic concept into a direction jointly bet on by capital and industry. But where the two lines intersect in the middle, it is still a blank slate.
To simultaneously possess realism, scalability, semantic annotation, and trainability, what is needed is not just an algorithm or a device, but a complete set of underlying capabilities from collection to processing to output.
This set of capabilities is precisely what Hesai has honed in the automotive supply chain over the past decade. Against this background, looking at Hesai's release of Kosmo, its uniqueness lies not in "yet another company doing 3D reconstruction," but in its identity as a physical AI infrastructure company.
For physical AI to be implemented, two problems need to be solved: how machines act on the real world, and how AI obtains sufficiently high-quality real-world data. Focusing on these two issues, Hesai has deployed two major engines.
One is the robot motion engine, corresponding to hardware-side capabilities. The core is the power module, used to drive robot joints and dexterous hands. A single humanoid robot needs to be equipped with more than 100 power modules, accounting for more than half of the BOM cost of the whole machine. Hesai's accumulation in precision machinery and transmission, chip design and algorithms, and servo control and sensing can be completely transferred to this field.
The other is the 3D world data engine, corresponding to data-side capabilities. Kosmo is its first systematic productized implementation. It consists of a handheld device integrating LiDAR, an 8K imaging system, and a high-precision IMU, along with a cloud platform. For ordinary indoor scenes, it takes about 15 minutes to complete a 100-square-meter space scan, and the model generation speed reaches five times that of similar products in the industry. But speed is only superficial; what is truly critical is Kosmo's structural design in data form.
Traditional 3D reconstruction outputs a "dead" model; the geometry is correct, the colors are correct, but the machine cannot see through the physical spatial positions when holding it. Kosmo's approach is to split a single reconstruction into three dimensions for simultaneous output: the visual dimension solves whether it looks realistic, the interactive dimension solves whether it can be touched or disassembled, and the spatial dimension solves how far away it is and how big it is.
The underlying layer of Kosmo World is Hesai's self-developed Kosmo Spatial Core, which synchronously outputs three forms of data in a single reconstruction: 3DGS is responsible for visual realism, restoring the appearance, materials, and lighting of the scene to solve the problem of "fake visuals"; mesh supports the splitting, modeling, and interaction of objects in the scene, allowing the robot to know that the sofa can be pushed away, solving the problem of "fake interaction"; depth provides precise spatial distance parameters, allowing the robot to judge how many centimeters the cup is from the edge of the table, solving the problem of "inaccurate space."
The three types of data originate from the same scene and are completely aligned in space; what the robot sees, grabs, and measures is the same world. In addition, AI semantic segmentation allows each object in the scene to be independently identified and annotated, and generative AI fills in unscanned blind spots. Kosmo can also perform visual adaptation according to the camera parameters of the target robot, allowing the robot to train in the world "in its own eyes," further narrowing the gap between simulation and the real machine. From collection to reconstruction, from reconstruction to semantics, and from semantics to adaptation, each layer is superimposed in the same set of coordinate systems, so the map truly changes from being viewable to trainable.
The motion engine allows machines to act in space, and the data engine allows machines to understand space. This set of capabilities is precisely the underlying barrier commonly needed by spatial data collection devices and robot power modules. So, with a real-world map and this combined capability, what can be done?
The value of a world map lies not in the moment it is drawn, but in someone taking it to navigate, to trade, and to explore unknown coastlines. How should this map made for AI be used?
Kosmo draws the map in a format that machines can directly read and train on. For the first time, the usability of the map is placed on an equal footing with "realism." And it is precisely this usability that gives the same map cross-domain spatial imagination; it is not just for robots to calculate distances, but also for artists to dream and for historical culture to archive.
In the field of embodied AI, spatial intelligence is the prerequisite for robots to understand the physical world, and the map drawn by Hesai Kosmo is the base map for robot actions. Taking a DQ ice cream store as an example, traditional manual modeling to restore a store takes about 300 hours. Kosmo compresses scanning and reconstruction to 15 minutes. After deployment, the robot's task success rate increases to 99%, and the failure rate drops to one-tenth of the original. The value here is not just speed, but that robots can repeatedly try and train in a truly restored space, without having to explore on the real machine time and again at the risk of damage. The cooperation between Yinhe General, Yuanli Lingji, and Hesai further turns real space into a scalable training asset.
In the field of film and television media, spatial intelligence becomes the migration of creative freedom, and Kosmo has also opened the doors to film and television giants such as Disney. Dylan Sisson, the creative lead of Pixar's RenderMan, used Kosmo to scan the iconic desk lamp in the park, completely preserving the texture of the floor tiles and the mottled reflections on the metal shell. The traditional process requires the team to set up heavy equipment, take tens of thousands of photos, and then have modelers spend weeks manually topologizing and texturing, while Kosmo compresses this step into a single scan. More importantly, Dylan turned the entire workflow into an official tutorial, open-sourcing it to creators worldwide. An animator can directly move the bookshelf in their home, the coffee shop on the corner, or an old house encountered during a trip into the digital world, turning it into a scene in an animation. Here, the map is no longer a navigation tool, but a film set that can be repeatedly framed and called upon. Here, the same map changes its face, rigor becomes a dream, and a single scan can dream infinitely.
In the field of cultural tourism and cultural heritage, spatial intelligence allows historical and cultural heritage to be permanently preserved in digital form. The cooperation between Hesai and Gaode has connected the first batch of Kosmo-reconstructed scenes to the "Flying Street View" of the Gaode APP, allowing users to freely browse 3D replicas of real spaces on their mobile phones. But the significance of this goes beyond "cloud shopping." After ancient buildings, old streets, and disappearing villages are completely recorded, they become an archive that will not weather. Cultural and museum institutions can use it for digital exhibitions, and designers can do secondary creations on it. Ordinary people can walk into a place they might never reach through a screen. Here, the map is both inheritance and the starting point of creativity. History is preserved, turning into a space that can be repeatedly opened, repeatedly told, and repeatedly recreated.
These three scenarios, one requiring rigor, one requiring dreams, and one requiring inheritance, seem scattered, but at the end of the day, what they require is the same capability: efficiently and high-fidelity transforming the real world into reusable 3D data. And precisely because this capability is fundamental enough, it can simultaneously support the computation of robots, the imagination of creators, and the memory of civilization.
It can be said that the infrastructure competition for physical AI has just begun. Fei-Fei Li says spatial intelligence is the next frontier of AI, and Jensen Huang says physical AI is the endgame of AI. This track has never lacked people who are optimistic about it; what it lacks are people who can turn blueprints into tools. Hesai has made the first move with Kosmo, drawing the world map in a format that machines can directly use. As for how far this map can ultimately extend, it remains to be seen how many robots, creators, and institutions are willing to set out with it.