SpatialView has published new work on using Radiance Fields as a data source for AI agents. By bridging reality capture and agentic workflows, the approach enables agents to analyze the physical world and extract actionable data directly from 3D reconstructions.
The work builds on SpatialView's earlier 3D search, which used text queries and original capture photos to find matching objects in Gaussian splatting models. While that method successfully localized objects, relying on original imagery presents a bottleneck for autonomous workflows. SpatialView notes that source photos aren't always available, and standard captures contain many redundant camera angles.
To address this, SpatialView's pipeline renders a fresh set of virtual photographs directly from the finished splat. The system evaluates candidate views for realism and selects a minimal set that maximizes total scene coverage.
Each view is then embedded with an image-text model and saved with the position it was taken from, forming what SpatialView calls a scene matrix. Asked a query like "where is the fire extinguisher?", the system finds the best-matching views and returns both the rendered view and its location in the 3D model.
On a 1.8-million-Gaussian scan of a cabin, with the source photos withheld, SpatialView reports that 122 virtual photographs were enough to index the scene, compared with 362 photos in the original capture, and framing and embedding took under two minutes.
SpatialView says keeping the index separate from the geometry means a newer embedding model can be adopted without touching the splat, and every answer comes with a rendered view a person can check. SpatialView says the precomputed index spares an agent the overhead of exploring the scene live for every query.
The write-up includes an interactive demo of the cabin that runs in the browser.
The full method is available now on the SpatialView blog.
