First in the world for 3D perception
on the Waymo open benchmark

The model Waymo submitted itself came thirteenth.

The category was 3D Semantic Segmentation on the Waymo Open Dataset, which measures how well a model tells one thing from another in a point cloud swept by a laser sensor.

We first learn the smallest geometric units in that cloud — points, edges, faces, holes — and then sit individual tasks such as weld seam recognition or progress measurement on top as a light layer. Adding a task does not mean retraining the whole model.