Odyssey, the world-model startup founded in 2023, unveiled Odyssey-3, a foundation model the company says can control robot arms, drive cars, pilot drones, and power humanoid robots from a single set of pretrained weights. Odyssey says the same system also learned to play commercial video games, including Rockstar Games’ GTA V, without game-specific training from scratch. The company plans a public release “in the coming weeks,” according to its own announcement.
The pitch is consolidation. Rather than building one narrow model per robot or per task, a single system would already understand how the physical world behaves and would need only a small amount of task-specific data to apply that knowledge. Odyssey describes Odyssey-3 as an autoregressive diffusion transformer trained on a large set of visual observations, from which it says the model acquired an implicit grasp of physics, cause and effect, and human behavior.
Odyssey’s own figures illustrate the sample-efficiency claim it is making. The company says Odyssey-3 learned to operate a range of robot arms from tens of hours of demonstration data, drove a car through closed-loop simulations of Indian city streets on 20 hours of simulated driving, and flew a drone around obstacles after training on tens of hours of simulated flight. On real roads, Odyssey says policies trained entirely in simulation managed about 77 percent of the distance that policies trained on real road footage managed before a safety driver had to step in.
Two robotics partners are named as validating the approach externally. Odyssey says it is working with Poke & Wiggle, which it describes as a robot-data and benchmarking firm, to test how the model’s knowledge transfers across different robot bodies and camera viewpoints. Separately, humanoid-robotics company Flexion built control policies on top of Odyssey-3 using what Odyssey describes as only tens of hours of teleoperation data; Flexion co-founder and CEO Nikita Rudin said in the announcement that the collaboration “opens up exciting possibilities for how quickly robots can acquire useful skills and adapt to unfamiliar situations.”
All of these results, the transfer rates, the hours-of-data figures, the game-to-game skill transfer, come from Odyssey’s own blog post. The company has not published independent benchmark comparisons or third-party replication of the driving, drone, or humanoid results, and the 77 percent simulation-to-road figure has no baseline model named for comparison. Odyssey itself flags that world models remain roughly two orders of magnitude smaller in scale than today’s language models, a gap it says is the reason world models at the frontier still trail the potential it describes.
The more notable claim is transfer across games rather than within one. A mobility policy given roughly two hours of GTA V footage went on to ride a horse in Red Dead Redemption 2 and a motorcycle in Square Enix’s Sleeping Dogs, Odyssey says, with no further policy training on either game. If that generalization holds up outside curated demos, it would matter more for robotics than for gaming: a warehouse-arm policy that adapts to a new gripper without fresh data collection, or a drone that handles an unfamiliar building layout, is the commercial case Odyssey and its rivals in world modeling, including Google DeepMind’s Genie line and World Labs, are all chasing.
For robotics teams evaluating foundation models this quarter, the question to press Odyssey on before the public release is reproducibility: whether the recovery behaviors and cross-game transfer hold on tasks the company did not select for the demo reel, and whether independent labs get access to run their own comparisons once Odyssey-3 ships.
Based on Odyssey’s own announcement, “Introducing Odyssey-3: A General-Purpose Physical Intelligence,” published on the company’s website.
