Odyssey
Odyssey · Video
Odyssey is an AI research lab whose foundation world models generate playable, interactive environments for robots, drones, games, and AI training.
About Odyssey
What Is Odyssey
Odyssey is an AI lab that builds world models: systems trained to predict what happens next in an environment, then render that prediction as video you can steer in real time. Instead of producing a fixed clip, the model keeps a running state of the scene and updates it every time you press a key or move a controller. The company was founded in 2023 by Oliver Cameron and Jeff Hawke, both veterans of autonomous driving, and has raised a $310M Series B.
The pitch is a shared foundation for physical and virtual systems. The same model family is used to drive robot arms, power humanoid control policies, fly drones, play games, and train other AI agents inside simulated worlds. According to Odyssey, its latest model, Odyssey-3, can pick up new robot tasks from tens of hours of demonstrations by leaning on what it learned from visual observations during pretraining.
The limits are worth knowing before you plan around it. Most of what Odyssey publishes are research previews rather than finished products, so access, quotas, and features change without notice. Long rollouts still drift, and the team says so: scene layout and acoustic identity can wander over time, and there's no standard benchmark yet for interactive audio-video generation. Visual polish also trails dedicated video generators, because the model spends its budget on responding to you rather than rendering a perfect shot.
Getting Started
- Open the Odyssey site and pick a demo: Odyssey-3 for single-player exploration, or Agora-2 for a shared multiplayer session.
- Load the experience in a desktop browser. The demos expect a keyboard, a gamepad, or on-screen controls on a phone.
- Move through the generated world and watch how the scene reacts. Each frame is produced on the fly, roughly every 40 milliseconds.
- For programmatic access, register on the developer platform and call Odyssey-2 through the Odyssey API to generate interactive video from prompts and input streams.
- Check the research posts before you build, since model names, preview limits, and supported inputs shift with each release.
Product Information
A quick look at Odyssey's pricing, supported platforms, and performance.
Best for
The users, tasks, and scenarios where this tool fits best.
Users
- Robotics and autonomy teams
- Game designers and prototypers
- AI researchers
Tasks
- Exploring a generated environment
- Training robot policies
- Stress-testing agents
- Prototyping game concepts
Scenarios
- A research sprint where you need an interactive environment this week
- Multiplayer experiments
- Sim-to-real prep for a robot arm
Key features
Real-Time Interactive Worlds
The core product is a world model that streams generated frames as you play. Odyssey-1 set the pattern by producing a frame roughly every 40 milliseconds, fast enough that keyboard, controller, or phone input feels connected to what you see. There's no render farm and no waiting for a clip to finish.
One Model Across Many Bodies
Odyssey-3 is an autoregressive diffusion transformer trained on visual observations of the world, and the same checkpoint is used for robot arms, humanoids, cars, drones, and games. The claim is sample efficiency: a new physical system needs hours of demonstrations, not a new dataset the size of a pretraining run. Odyssey reports recovery behaviors, such as regrasping after a missed pick, that never appeared in the training demos.
Multi-Agent Shared Simulations
Agora-2 puts several participants, human or AI, inside one generated world at the same time. The architecture splits the job in two: one model evolves a shared world state from everyone's actions, a second renders each viewpoint consistently. Up to four players can share a session, which is the closest thing yet to a game engine where both halves are learned.
Audio and Video Together
Starchild-1 generates sound and picture in the same rollout, at up to 24 fps, and keeps them in sync through an asynchronous cache design that lets each modality run on its own clock. It responds to streaming text, speech, or action input, and supports scripted dialogue, conversation, and narration modes. Sound matters for physics and intent, not just atmosphere.
World Models as Training Grounds
PROWL-2 closes the loop between an agent team and its world. Each side runs its own curriculum inside a single training loop, so the environment gets harder as the agents improve and the agents get data from their own failures. For teams training embodied agents, that replaces a hand-built simulator with one that grows with the policy.
Browser Demos and a Developer API
You can try the models without writing code: experience.odyssey.systems hosts the Odyssey-3 demo and agora.odyssey.systems hosts the multiplayer one. For production work, Odyssey-2 is available through the developer platform, where you send prompts and input streams and receive interactive video back.
Pros and cons
Pros
- Interaction is the point, not a feature bolted onto a clip generator, so input changes the next frame instead of the next render.
- A single foundation model covers robots, drones, vehicles, and games, which cuts the number of systems a team has to maintain.
- Data efficiency is real enough to matter: tens of hours of demonstrations can produce working robot control in Odyssey's reported experiments.
- The research is public, with technical reports, papers, and playable demos you can test before committing.
- Multi-agent support makes shared simulated worlds possible without hand-authoring a map.
Cons
- Most releases are research previews, so availability and limits can change between one project and the next.
- Long sessions drift: scene layout and audio identity wander over time, and physical accuracy still varies across robot bodies, so long continuous shots and safety-critical work need checks.
- There's no published price list. API access and enterprise work go through the developer platform or sales, so budgeting takes a conversation.
- Visual fidelity sits below dedicated video models, and you give up exact control over geometry that a hand-authored engine would give you.
Frequently asked questions
Odyssey builds world models that generate interactive video in real time. You prompt a scene, then control it with a keyboard, gamepad, or phone, and the model predicts and renders what happens next.
Related content
Explore related tools, skills, and articles for Odyssey.
Odyssey Alternatives
Griffin by Tavus
Tavus · Voice & Language · Video · ChatbotGriffin by Tavus is the first Human Interaction Model: a full-duplex, video-to-video system that watches, listens, and answers during a live video call in real time.
Syllaby AI Avatar 2.0
Syllaby · Video · MarketingSyllaby AI Avatar 2.0 turns a script into a presenter-led video with a reusable AI avatar, natural voice, B-roll and automatic subtitles, all in one browser workflow.
Vitra.ai
Vitra.ai · Voice & Language · Video · MarketingVitra.ai is an agentic content platform that creates, translates, reviews, and publishes video, images, documents, and app content in 120 languages from one shared brief.