RESEARCH DIRECTION
Language models are eloquent but unsituated — fluent describers of a world they have never inhabited. We hold that spatial understanding is not an application of intelligence but a constituent of it: representing scenes with geometry and physics, simulating how they evolve, and closing the loop between perceiving a world and acting within it.
Text is a lossy projection of reality. Systems trained only on that projection inherit its gaps — no occlusion, no object permanence, no contact dynamics. We work on representations that recover what language discards: metric structure, spatial relations, and the physical constraints that make some described worlds possible and others nonsense.
A spatially intelligent system should be able to reconstruct a scene from partial observation, generate coherent extensions of it, and simulate its evolution under intervention — three capabilities that share one underlying world representation. Multimodal world models that unify these are, in our view, a load-bearing component of general intelligence.
Embodiment is the forcing function for grounded semantics: an agent that must act on its representations discovers immediately which of them are wrong. We study closed perception-action loops — real and simulated — as both a training signal and the most honest evaluation an intelligent system can face.
WORKING PRINCIPLES
CONTINUE EXPLORING