Odyssey introduces Odyssey-3, a preview world model for physical control

• Odyssey has previewed one foundation world model for robot arms, humanoids, cars, drones, agent-training environments, and games, as testingcatalog's announcement describes.
• Task adaptation keeps the world model fixed and trains a smaller action decoder for each control interface, according to rohanpaul_ai's breakdown.
• The driving demonstration used 20 hours of simulated data, rohanpaul_ai's driving note says; Odyssey reports that its simulation-only policy traveled about 77% as far between safety-driver interventions as a policy trained on real footage.
• Odyssey calls Odyssey-3 an early preview, with a public release planned within weeks, per testingcatalog's follow-up.

Odyssey says in its launch post that a policy trained on roughly two hours of GTA V footage produced horseback movement in Red Dead Redemption 2 without training on that game. The same post puts its simulation-trained driving policy at roughly 77% of the intervention-free distance of its real-footage counterpart.

Odyssey calls the core model an autoregressive diffusion transformer trained on broad visual observations of the world. Its published training recipe separates the shared world representation from each machine's controls:

• The pretrained world model remains frozen during downstream training.
• Experiential data pairs the system's observations with actions taken for a task.
• A learned action decoder, or small task policy, maps Odyssey-3's internal visual representations to the target robot, vehicle, drone, or software controls.

Odyssey reports the following results in its launch post:

• Robot arms: Tens of hours of demonstrations supported manipulation tasks. Odyssey says it observed recovery after missed grasps and unusually positioned dropped objects.
• Humanoids: Flexion built whole-body policies from tens of hours of teleoperation data. Odyssey says its tested policies continued under lighting changes that caused its VLA baselines to fail.
• Driving: Odyssey says a policy trained on 20 hours of simulation drove closed-loop on Indian streets. The setup rohanpaul_ai highlighted used simulated experience only, while Odyssey reports about 77% of the intervention-free distance achieved by a policy trained on real driving footage.
• Drones: Tens of simulated flight hours trained an indoor navigation policy that generated waypoints from camera views, drone state, and a high-level prompt.
• Games: Policies trained from recordings paired with keyboard and mouse inputs produced extended GTA V sessions. Odyssey reports that a mobility policy trained on about two hours of GTA footage transferred to horseback movement in Red Dead Redemption 2.

Odyssey also positions Odyssey-3 as a generated environment in which other agents can take actions and learn their consequences, a use case testingcatalog included in its overview. Its PROWL work uses agent-discovered failures to guide world-model improvements, while more reliable simulated experiences become training data for the agents.

Odyssey's current API documentation introduces Odyssey-2 Pro, with interactive-stream, viewable-stream, and simulation endpoints. The Odyssey-3 announcement promises a public release in the coming weeks rather than naming an API, model-weight, or pricing path.