Technology

How a robot
builds a map.1

It starts as a rough guess.

Gaussians move, resize and recolour, again and again.

Until the scene matches what the cameras saw.

Gaussian Splatting

A scene made of small, soft, coloured blobs called Gaussians. Seen together, from any angle, they look like a photograph of the place.

Building one takes a datacenter GPU.
Mapping should happen on the robot.

Today1

Robots record now and get their map hours later, from a server.

The goal1

The map grows on the robot, while it moves.

What already runs

This frame was rendered by our hardware.2

Lego bulldozer scene rendered on our FPGA prototype
173,161Gaussians in the scene
1080pend to end
262 msper frame, down from 1.02 s in 10 days
Concept render of the Ateno AX chip in its package

Ateno AX

A chip built for one job:
3D mapping on the device.3

Builds the map.

Training in hardware, designed for edge power budgets.

Renders it.

A full pipeline, from Gaussians to pixels.

Fits a robot.

As a compute module or a licensed block.

Concept render of a black Ateno AX module connected to a laptop by a USB-C cableConcept render close-up of a USB-C plug seated in the Ateno AX module

Ateno AX

Plug it into your computer.3

We're designing Ateno AX as a module that connects to your computer over USB-C.

Datacenter GPU and Ateno AX

The same job.
A very different place to do it.4

Illustration of a datacenter aisle lined with server racks

Datacenter GPU

Hours later.

Concept render of a warehouse robot carrying an Ateno AX module

The goal with Ateno AX

While it moves.

ComparedDatacenter GPUAteno AX goal
Where training runsIn a datacenterOn the robot
When the map arrivesHours laterWhile it moves
Power it has to fitDatacenter power and coolingAn edge power budget
What it depends onA link to a serverNothing off board

Let's talk.

Book 30 minutes with the founders.

Book a call