Oasis generates simulated environments for training robots and self-driving systems
Decart's other model builds simulated worlds intended for training and testing robotics and autonomous driving. Generated environments are how you get the millions of hours of varied experience that physical robots cannot accumulate on a schedule.
Embodied learning has a data problem that language does not: there is no internet of robot experience to scrape. Every hour of manipulation or driving data costs an hour of real hardware in the real world, and rare events — the ones that matter — are rare by definition.
Generated environments attack that directly. A world model that produces varied, physically plausible scenes on demand turns data collection into a compute problem, and compute scales in ways that fleets of physical robots do not.
The standing objection is the sim-to-real gap, and it is real. A policy trained in a generated world learns that world's physics, including its errors, and transfer has historically been the point where impressive simulation results stop being impressive. Nobody has retired that concern.
But the demand side is unambiguous. Two thousand robotaxis heading into five European cities need validation against conditions that have not happened yet, in cities where the fleet has no history. Generated environments are one of the few ways to test for events you cannot wait for.
Quartz — Anthropic in talks to acquire Israeli AI startup Decart for $6 billion → · citybiz — Anthropic Reportedly in Talks to Acquire AI Startup Decart in $6 Billion Deal →