Hesai launches a camera for turning real rooms into robot training worlds

Hesai launched Kosmo on 14 October, pairing the Kosmo One camera with Kosmo World cloud software to reconstruct scenes for robot simulation and visual production.[1]

The camera synchronises lidar, imagery and motion measurements. Hesai says a typical 100-square-metre indoor space takes about 15 minutes to scan, depending on layout.[1]

Hesai specifies 8K capture but export up to 6K. Output options include Gaussian splats, meshes and depth data, with support announced for NVIDIA Isaac Sim and Isaac Lab. These are supplier specifications, not independent test results.[1]

A scene is more than a photograph

The original 2023 Gaussian-splatting research helps explain one of those outputs. Instead of describing a scene only as solid surfaces, the method represents it using many three-dimensional Gaussian elements—soft, spatially extended points whose combined appearance is optimised against photographs. It can then render views from different camera positions.[2]

That makes a useful distinction visible. A conventional photograph fixes the viewpoint. A reconstructed scene aims to let a virtual camera move through the space while preserving how it looks. Gaussian splatting is a way of producing those views efficiently; a geometric mesh is a different description, built from connected surfaces.[2]

Looking right and behaving right

Simulation requires another layer. NVIDIA’s Isaac Sim tutorial shows an object with rigid-body dynamics falling through a floor until collision properties are added. It separately configures mass and surface behaviour, including friction and restitution—the tendency to bounce. A convincing picture therefore does not automatically tell a simulator how objects should interact.[3]

Consider a hypothetical robot learning to collect a mug from a shelf. The visual scene can help it recognise the mug and locate the shelf. Testing a grasp also needs assumptions about contact, mass and how easily the mug slips. Moving the shelf or changing the lighting would create further training cases, but changing the assumed friction tests a different failure.

The software can also generate missing parts of a scene. Such completion fills a digital environment; it does not document what an unscanned area actually contained.[1]

The appeal is a shorter route from a real workspace to a reusable digital one. Our reading is that capture quality and simulation quality remain complementary tasks: first acquire the room, then specify the behaviour the robot must learn inside it.

Report an error or suggest a correction