technology-hero
Technology

3D perception that understands the site,
technology that moves the robot

We work out where the robot stands, recognise precisely what it has to work on, and carry sensor data through to the robot system itself.

Approach

Flexibility and precision,
solved in separate layers

One layer reads the site, one adapts to a site it has never seen, and one does the work.
Change a layer and the others stay as they are.

Perception We read the site in 3D first The model first learns the smallest geometric units in the point cloud — points, edges, faces, holes. Weld seam recognition or progress measurement then sits on top as a light layer. We call it the LiOps Spatial Model.
Adaptation Ten photographs are enough No retraining per site. Ten labelled photographs adapt the model to a new floor, and matching design drawings against what is actually there is handled on its own — at 100% on structures still under construction.
Execution Cumulative error within 1 mm Error accumulates at every hop between the coordinate frames of sensor, drawing, robot and tool. We hold the whole chain within 1 mm. Precision comes from software, so a single sensor under KRW 500,000 reaches the same figure.
01 · LOCALIZATION

Start away from the dock,
hold position as the site changes

Knowing where the robot stands is where every task begins. We match the prior map against what the sensors see right now to recover the initial pose, then carry straight on into continuous tracking.

Map registration is coupled directly into the estimate, so five kilometres in, the position has not slipped. Stock moves and the sensor is briefly blocked, and the estimate holds.

Read more →
Driving through unmapped ground while the pose estimate holds
The plate and its joint, reconstructed from stereo images
02 · STEREO DEPTH

Half a millimetre,
from stereo cameras alone

A model trained on the disparity between two views rebuilds the depth and shape of the workpiece. Inference on site runs on stereo images alone.

In the coil workpiece evaluation it held a mean error of 0.5–0.6 mm against Alsontech measurements. It reads gaps, steps and curved surfaces, then hands the weld datum and position correction to the robot.

Read more →
03 · ROBOT SYSTEM

Sensors, algorithms and remote
visualisation on one stack

Robot System is our own robot platform. It carries sensor capture, delivery between algorithms, frame management, remote visualisation, recording and replay on a single stack.

Within one machine, shared memory cuts the cost of moving large images and point clouds; between machines, remote transport carries the floor's data. It sits on top of what you already run.

Read more →
Pick-up and drop-off zones laid over a plant point cloud, seen remotely
Proven

Already first
in the world

First worldwide on the Waymo open benchmark

On the public leaderboard for 3D perception in autonomous driving, our co-CEO Seungyeop Sun placed first. The model Waymo submitted itself came thirteenth.

Waymo Open Dataset · 3D Semantic Segmentation · 2022

Two patents granted for real-time 3D registration

Granted for registration and for assembly progress evaluation. Six more are filed and under examination.

KR 10-2983357 · KR 10-2893463

First in Korea to pass 1 mm in a government programme

Free-form butt welding, developed jointly with HD Hyundai Samho, was verified by a third party.

March 2026
Measured

The numbers do the talking

0%
Productivity gain
Measured on a site running Spatial MES.
0%
Drawing alignment success
Includes structures still under construction and not yet complete.
0mm
Cumulative error
Held within 1 mm across every coordinate frame: sensor, drawing, robot and tool.
0 photos
Adapting to a new site
Ten labelled photographs, demonstrated on a forklift task.
0cm
Localization error
Held within about 2 cm even over runs of 5 km and longer. From an IsaacSim stress test.
0mm
Stereo depth accuracy
Mean error of 0.5–0.6 mm against Alsontech measurements in the coil workpiece evaluation.

Data stays on premises. Nothing is sent to an external cloud.

Integration

We speak
your stack

We do not make you conform to our specification. The data and commands you need are delivered in the operating system and the language you already run. Windows or Linux, and any perception software you already own can sit on top unchanged.

  • Interface specification
  • SDK and per-language clients
  • Example code and a mock server
  • Test tooling and a data dictionary
  • Version compatibility policy
Data flow background

Find out with LiOps
whether it works on your floor

Send us your drawings and site conditions and we will come back with whether it applies and what to expect.

Request a quoteTalk to us