Hesai supplies lidar observations for robots, vehicles and physical AI

Understand Hesai’s JT128 and ATX lidar, development resources, commercial scope and the difference between a sensor and an autonomous application. This is a public-source assessment with a proposed evaluation, not a hands-on test.

01 / ProductHesai sits at the observation layer of physical AI

Hesai makes lidar sensors used to observe three-dimensional surroundings. The current portfolio addresses driver assistance, autonomous mobility, robotics and industrial sensing. Its connection to AI is concrete: these measurements can supply the spatial input used by perception and navigation systems. Buying the sensor does not automatically provide the complete software or machinery that acts on those observations.

Two products show the difference between sensing jobs. JT128 is designed for robotics and industrial applications, with a published 360-degree horizontal and 189-degree vertical field of view. ATX is a compact forward-looking automotive product. The choice is not a simple progression from smaller to better; a robot observing nearby surroundings and a vehicle observing a road ahead have different requirements.

A Hesai announcement about automated vehicle marshalling explains a useful division of responsibility. Hesai supplies sensors, Outsight supplies perception software and Embotech supplies the system used in BMW facilities. That is vendor-reported deployment evidence and a reason for editorial inclusion. It also shows why a sensor supplier's role should not be confused with ownership of the whole autonomous workflow.

02 / AudienceRobotics developers need a view that fits the machine

Hesai is relevant to a team building a mobile robot, a vehicle platform or an industrial spatial-sensing application. Such a team needs to decide what data it will consume and who will turn it into operational behavior. The right starting point is the required observation: nearby obstacles, navigable space, a distant object or the movement of equipment through a fixed area.

A warehouse automation buyer who wants a complete supported transport process should begin with the system integrator or robot supplier. Selecting a lidar independently can be useful for a development organization, but it can also create an integration responsibility that an operator does not intend to own. Establish the boundary between component evaluation and application procurement early.

The Teradyne blueprint provides a contrast through robot arms and mobile material-handling products. The NVIDIA blueprint covers compute and software that may surround a perception pipeline. These are different layers of an automation project. A Hesai sensor can be valuable within a larger architecture without replacing the products and engineering work at those other layers.

03 / WorkflowA proposed robot evaluation should preserve the path from packet to obstacle

Consider a development team investigating a compact warehouse robot's ability to observe low obstacles and nearby overhangs. Start by mapping the parts of the machine that can obscure a sensor. The proposed evaluation should include the robot's normal loads and attachments, because a clear view on an empty chassis may not survive the actual operating configuration.

Choose the sensor with the required geometry in mind. JT128's broad vertical coverage is relevant to simultaneous observation of nearby ground and overhead space, according to its product page. Treat that as a reason to test the installation, not proof that there are no blind spots after mounting. The robot's body and the surrounding scene remain part of the optical problem.

Use the Hesai downloads page to identify the exact product manual, calibration material and available software. The page provides product-specific documentation and PandarView visualization tools; the developer portal provides a separate technical entry point. Record the device, firmware and software combination used in the trial so later observations can be reproduced.

Begin with stationary captures of known objects. Check that the expected geometry appears consistently and that the data is being interpreted in the correct coordinate frame. A rotated or displaced point cloud can look plausible while being unusable for a robot's local map. Verify the relationship between the sensor frame and the machine before judging a navigation algorithm.

Then introduce movement at controlled speeds. Include narrow objects, low features, changes in floor level and overhanging structures that resemble the proposed workplace. Retain raw observations and the interpreted obstacle output for the same interval. If the robot's planner misses an obstacle, those records help distinguish sensing, data conversion, perception and planning failures.

Test situations where observations become incomplete. An occluded surface, an object close to the body or a changed mounting angle should have a defined consequence for the application. The important question is how the robot behaves when evidence is insufficient. A colorful map is not an acceptance criterion for continuing motion through uncertainty.

Finally, compare the observation quality with the application's practical requirement. If the machine moves slowly in short aisles, a distant maximum-range number may be less valuable than dependable nearby coverage. If the job is a forward vehicle application, repeat the evaluation with appropriate target distances and surfaces instead of reusing the warehouse result. Product families and operating conditions should not be blended into one performance claim.

04 / PricingQuotes need hardware, support and integration boundaries

The JT128 and ATX pages direct buyers to contact sales or distributors. They do not establish a universal numeric price for a production unit or a complete robot installation. A useful quote identifies the exact hardware, quantity, accessories and intended support route. Public specifications alone do not establish regional inventory or a delivery commitment.

Hesai's published warranty policy, dated 31 March 2022 and still linked from the current site, states a one-year warranty from delivery unless the applicable sales contract agrees a longer period. It describes mail-in service unless otherwise agreed and says shipping expenses are negotiated. These are the published default terms, not a promise that every new commercial contract has identical service arrangements.

For a robot expected to operate continuously, replacement logistics can matter as much as the repair entitlement. Ask whether the project needs spare sensors, who performs replacement and how calibration is restored. Keep those application responsibilities separate from the manufacturer's warranty. A working replacement component does not by itself revalidate the robot's complete behavior.

05 / DistinctionsA broad sensor portfolio lets teams match geometry to the task

Hesai's strongest relevance is its focus on the sensing layer across several physical applications. The current site exposes product families with different fields of view and ranges, alongside manuals and development resources. In our assessment, this makes the company useful to teams that want to choose and integrate a specific observation capability rather than buy a general AI service.

ATX's current specifications list 230 meters at 10% reflectivity and a 120-by-20-degree field of view. The same page cautions that published values are typical under specified test conditions and may vary with hardware, firmware and application conditions. Those qualifications belong beside the numbers. An engineer should obtain the agreed specification for the actual variant before freezing a vehicle design.

The marshalling example also demonstrates a different architecture from onboard autonomy: sensing can be part of fixed infrastructure that supports a larger operating system. This broadens the set of reader questions Hesai can help answer. It does not mean the same sensor selection or software stack fits every factory, robot or vehicle.

06 / QuestionsProduct variants and data quality deserve more attention than maximum range

JT128's specification table distinguishes 40 meters at 10% reflectivity from a 60-meter maximum. That difference is material when a buyer's important obstacle is dark or difficult to observe. Define the target and required detection behavior before using a range figure in a comparison. A maximum distance without target conditions is a poor proxy for the application's useful perception envelope.

The JT128 page also covers JT64P, while its copy mentions capabilities of the broader platform. Avoid combining the channel count, output rate or performance of different variants into one imaginary configuration. Put the full product designation on each test record and purchase line so a successful evaluation remains tied to the device actually ordered.

The current homepage advertises Kosmo, an AI spatial camera, through an early-access application. That is a different availability status from a blanket general-release claim. This blueprint's proposed workflow stays with the documented lidar development path; it does not assume a reader can immediately purchase or deploy every product shown on the homepage.

No hands-on performance or regional purchasing eligibility was verified for this review. For a real project, establish the delivery route and supported deployment conditions with the supplier and integrator. The most useful next evidence is a representative recording and a clear account of how the application behaves when that recording is incomplete.

07 / DecisionEvaluate the sensor inside the system that will use it

Hesai belongs on a shortlist when a physical AI project needs defined spatial observations and the team can own or commission integration. Match the sensor's geometry to the task, preserve the software and hardware configuration, and measure the downstream result. The decision becomes stronger when the buyer can explain precisely what the sensor contributes and what remains outside its scope.