Odyssey introduced Odyssey-3 on October 8 as what the company calls its most powerful foundation world model yet. In its launch post, Odyssey said the model is meant to generate interactive experiences, train AI systems and power robots. The company made a research preview available for free on launch day.
The pitch goes beyond generating a finished video. A demo description says Odyssey-3 builds an environment from a prompt and keeps producing the scene as a person moves through it. Walking, jumping or moving the camera can change what appears next, with examples including ripples after entering water and a car crossing a street.
A model that has to react
A world model attempts to represent how a scene changes after an action, rather than only rendering a predetermined clip. That makes consistency and physical behavior central to the release.
Odyssey claims the model set a new state of the art on Physics-IQ. One post explaining the benchmark described it as a test in which models continue videos of real physics experiments. Another launch summary reported that Odyssey-3 Pro scored 66.1 on the verified video-to-video test using best-of-eight sampling.
From virtual scenes to machines
The same launch summary said Odyssey adapted the model to robot arms, humanoids and a car driven on roads in India, with a separately trained policy for each system. The benchmark explainer also reported that a robot arm recovered from a missed grasp even though that behavior was not present in tens of hours of demonstrations.
Odyssey-3's research preview and physical-system adaptations point to the wider goal behind the release: using a learned visual model to predict what should happen next in both simulated environments and machines.
Comments