Hesai's Kosmo Wants to Be the Camera That Teaches Humanoid Robots to See

The most revealing product launch in robotics this week did not come from a robot company at all. Hesai Group, the Chinese lidar maker whose sensors sit on Mercedes and Ford driver-assistance systems, unveiled a device on Wednesday that looks like a wine bottle and behaves like a camera. It is called Kosmo, and its job is to feed the one thing humanoid robots are starving for: real-world spatial data.

David Li, Hesai's co-founder and CEO, introduced Kosmo at the company's Shanghai factory with a blunt diagnosis of the industry he is now courting. There is, he said, a significant gap in the high-quality spatial data needed to train AI-powered robots to understand and interact with the real world, and to perform complex tasks that go beyond waving arms or dancing. The line landed because it was honest: most of today's humanoids remain confined to low-impact functions like greeting guests, waving, or dancing, in Li's own words. The videos are great. The productivity is not.

What Kosmo actually is

Kosmo combines lidar sensors and cameras in a single unit to capture three-dimensional maps of physical spaces. The point is efficiency. Developers training robots need dense, accurate spatial data of factories, stores, warehouses, and streets, and collecting it has historically meant strapping together separate lidar units, cameras, and synchronization hardware, then processing the results by hand.

Hesai's pitch is that Kosmo makes that pipeline cheap and fast enough to become routine, the way phone cameras made collecting images routine a decade ago. If robots learn by observing the physical world, the company selling the camera is betting on the chokepoint of the whole field. Li declined to give revenue projections for the product but said the company has signed agreements with Chinese robotics firms to use it.

The data bottleneck is real

This is worth taking seriously even coming from a vendor, because the bottleneck Li describes is one of the least controversial claims in robotics research. Modern robot brains are trained the way large language models are: on data. But while the internet handed AI companies trillions of words for free, there is no internet of touch, depth, and physical interaction to scrape. Every millimeter of useful three-dimensional understanding has to be collected deliberately, in real spaces, with real hardware.

That is why the autonomous driving industry became, in effect, the robotics industry's data-collection apprenticeship. Fleets of lidar-equipped cars spent years mapping streets, and Hesai rode that wave to become, by Reuters' account, the world's largest lidar supplier for autonomous vehicles over the past three years, driving the cost of the technology below $200 per unit. Now the company is trying to run the same play again, except this time the fleet is any developer holding a Kosmo, and the streets are warehouses, restaurants, and factory floors.

"What we aim to offer is essentially a platform for the underlying technologies required by the robotics industry, positioning Hesai as a tier 1 supplier comparable to Bosch in the automotive sector, or perhaps even beyond." (David Li, CEO of Hesai)

The Bosch ambition, and the actuator half of the plan

The Bosch comparison is telling. Bosch does not sell cars. It sells the parts, systems, and software platforms that nearly every carmaker depends on, and it is enormous precisely because it is invisible. Li is saying out loud what many component makers believe quietly: the money in the humanoid era may not be in building the robot, but in selling the senses, joints, and data plumbing that every robot needs.

Hesai is already moving on the second half of that plan. The company is developing robot actuators, the devices that convert electricity into mechanical motion, which Li says account for more than half of a robot's material costs. Demand is growing, he added, for actuators that are lighter, smaller, and more precise, so robots can become safer and more agile. If lidar taught Hesai how to drive costs down through scale, actuators are the obvious next mountain: they are the most expensive line item in the machine, and whoever commoditizes them shapes every humanoid's price tag.

The Dairy Queen test case

The most concrete evidence Hesai offered is already in a Shanghai restaurant. The company's actuators are being used by Sharpa Robotics, a startup founded by Hesai's co-founders, in its flagship robotic hand, the Sharpa Wave. A Sharpa Wave robot has been autonomously serving ice cream at a Dairy Queen outlet in Shanghai, and Hesai said Kosmo's detailed mapping technology helped the robot achieve a task success rate of 99 percent.

Read that with the usual vendor-claim caution. Sharpa operates independently of Hesai, though it functions as a testing ground for Hesai's robotics technologies, and the deployment is one of the industry's first attempts at complex production work in a live commercial environment. At the chain's Wujiang Road outlet in Shanghai, Sharpa and Dairy Queen opened what they described as the world's first restaurant capable of year-round autonomous robotic operation without modifying the existing store. Independent coverage has noted the deployment has been characterized at times as teleoperated rather than fully autonomous, so the 99 percent figure belongs to Hesai, not to an independent audit. Still, a robot making frozen treats through dozens of consecutive manipulation steps in a store built for humans is a more serious benchmark than a trade-show dance routine.

Sharpa itself is not short of capital. The company recently raised more than 4.5 billion yuan, about $672.4 million, in a funding round backed by Alibaba, Tencent, JD.com, Meituan, and others, at a post-money valuation of 22 billion yuan, roughly $3.3 billion. It debuted its North robot at CES in January, unveiled a new hardware lineup at IROS in September, and has said it plans to work through catering, hotel, and retail deployments before attempting household products from 2028.

What to watch

Kosmo matters less as a gadget than as a strategy. Hesai is trying to own the boring, expensive, unavoidable parts of the humanoid stack. First it drove lidar below $200 a unit for cars. Now it wants to sell the camera that maps the physical world for robots, and the actuators that move them. If that works, Hesai becomes the supplier that every humanoid company must call, regardless of which robot wins the hype cycle.

The risks are the familiar ones. Agreements with Chinese robotics firms are not the same as an order book, and the spatial-data bottleneck exists partly because real environments are messy in ways a better camera cannot fully fix: lighting changes, occlusions, things that move. And the Dairy Queen milestone, impressive as it is, is one restaurant doing one task.

But step back and the pattern is hard to dismiss. The last decade of AI taught us that data beats cleverness: the model with the most training data usually wins. Humanoid robotics is about to run the same experiment with physical data. Hesai just sold the industry a shovel, and history says the shovel sellers rarely do badly.