Odyssey Launches Odyssey-3 World Model for Embodied AI

Odyssey has introduced Odyssey-3, an autoregressive diffusion transformer designed to act as a shared foundation world model for embodied AI. Instead of training separate models for different hardware, developers can keep the core Odyssey-3 backbone frozen and train compact, lightweight action decoders for specific systems. This architecture allows the same underlying model to translate observations into commands for robot arms, humanoids, autonomous vehicles, drones, and video game characters.

In early tests, Odyssey reported notable cross-domain capabilities. A driving policy trained entirely in simulation achieved 77 percent of the distance between safety-driver interventions on Indian roads compared to policies trained on real-world footage. In humanoid robotics tests conducted with Flexion, the model successfully generalized to lighting changes that caused standard vision-language-action baselines to fail. Additionally, a mobility policy trained on Grand Theft Auto V demonstrated zero-shot transfer by successfully controlling horseback movement in Red Dead Redemption 2 without any game-specific training data.

For practitioners, this approach could significantly reduce the time and data required to deploy AI on new physical platforms. Instead of collecting massive datasets for every new robot body, developers only need to train a small decoder. Odyssey is also utilizing the model in a companion project called PROWL, which serves as an interactive simulator where AI agents can train and learn from their mistakes in a closed feedback loop.

Although Odyssey has partnered with Poke & Wiggle to benchmark the model across different bodies and viewpoints, independent validation is still pending. The company has not yet disclosed details regarding parameter counts, pricing, or hardware requirements, but it plans to release Odyssey-3 publicly in the coming weeks. The model is already deployed with select partners in the robotics, autonomous driving, gaming, and defense sectors.