Speridlabs comes out of stealth to build a 'Spatial AI' in 3D

Speridlabs exits stealth with backing from Pear VC and Base10 to build 3D Spatial AI foundation models, aiming to solve data scarcity and hardware costs.

Speridlabs, a research lab backed by Pear VC and Base10, is publicly launching with a precise ambition: to design foundation models capable of understanding the world in three dimensions, to perceive, reason, simulate, and generate within it. The team begins with a critical observation. Computer vision, in their view, remains a field conceived in 2D, fragmented into isolated tasks such as detection, segmentation, or reconstruction, whereas most of these problems are actually spatial. To illustrate this limitation, the lab points out that an autonomous driving software stack currently assembles more than forty distinct models, stitched together by heuristics and human supervision.

The project relies on a unique model that would maintain a consistent 3D representation regardless of viewpoint, occlusion, or elapsed time. Reconstruction, reasoning, editing, navigation, or generation would thus become so many queries addressed to a single base. Four obstacles are identified: data scarcity, with approximately 100,000 public 3D scenes compared to billions of images for current models, the absence of a standard representation for 3D, evaluation protocols inherited from 2D, and the hardware cost of spatial processing.

The roadmap is progressive: first, reconstruction priors derived from real-world captures, then a 3D generative model also serving to produce data, and finally a world model that is initially static then dynamic. Speridlabs indicates its intention to publish its work, benchmarks, and model weights, due to a lack of shared knowledge on the internal workings of these systems.