Niantic's Physical AI Spinoff Takes on the Real World, One City at a Time
/Niantic Spatial’s CEO contrasts the real-world margin for error against those we’ve grown accustomed to dealing with from LLMs and the like.
By way of explaining why simulation has become such an important aspect of robotics training, the company notes in a recent blog post, “Running [reinforcement learning] in the real world is hard, because robots do not get many second chances. A misjudged gap or a collision with a glass door is expensive, slow to reset, and can damage hardware.”
(There’s also the matter of humans, animals, and vehicles to contend with — but let’s not get ahead of ourselves.)
Niantic Spatial’s core proposition isn’t world models, so much as models of the world — large geospatial models (LGM). It’s a kind of digital twin for reality, built on the supposition that the best way to accurately train robots in simulation is to hew as close to the real world as is plausible.