Odyssey-3 — One world model for robots, cars, drones, games
On September 15, 2026, Odyssey unveiled Odyssey-3, a foundation world model (an autoregressive diffusion transformer) trained on visual observations of physics, dynamics, and human behavior. The same model is claimed to control robot arms (tens of hours of demos for pouring, packing, box-closing), power Flexion humanoids in real time, drive closed-loop on Indian streets from 20 hours of simulated data (~77% as far between interventions as real-footage policies), pilot drones, generate environments for recursive AI training (PROWL), and play games such as GTA V with zero-shot transfer to other titles. Collaborations include Poke & Wiggle and Flexion. It is an early bid for a shared physical-intelligence model that adapts across robots, vehicles, and simulations with little extra data.
x.com