Niantic Spatial Places Library

Niantic Spatial curated the places, and its embodied AI team announced them.¹ Each is one USDZ file holding two things derived from the same capture.¹ One is a photorealistic Gaussian splat, a way of representing a captured scene as many small coloured, semi-transparent blobs. The other is a collision mesh aligned to it: the plain surface geometry a physics engine uses.¹ Both are gravity-aligned and at metric scale.¹ The first catalogue covers industrial, logistics, commercial and residential settings, and Niantic says it will grow on request.¹

Niantic calls the intended use real-to-sim-to-real.¹ A team reconstructs the site a robot will work in. It then trains and evaluates the robot's policy (the learned rules that map what it senses to what it does) there, in NVIDIA Isaac Sim, Isaac Lab or another OpenUSD simulator. The team finds where the policy fails and feeds that back.¹ Real places are also offered as a base for domain randomisation, which varies a scene's details so a policy does not learn only one setting.¹ Selected scenes can be relit to produce lighting and weather variants, a feature in beta.¹

Access is through Niantic Spatial's embodied AI plans, which also include capture minutes for reconstructing a customer's own sites.¹ The number of places depends on the plan, and one sample scene is free.¹ Prices and licence terms are not published.¹ The library was announced on 25 September 2026.¹