Odyssey-3: how a video model teaches robots physics in tens of hours

What is a world model

A world model predicts the next state of the physical world: it takes a video frame and an action (a command, a hit, a push) and generates what happens next. Odyssey-3 is an attempt to make such a model general-purpose and efficient enough to control real systems.

Odyssey was founded in 2023 by Oliver Cameron and Jeff Hawke, who spent a decade working on self-driving cars. Their reasoning is straightforward: a car needs to predict its surroundings. They watched robotics move through narrow, single-task systems, each requiring huge volumes of task-specific data. A human, by contrast, can sit down at a dangerous piece of machinery at 18 and learn it quickly, because they've spent their whole life watching objects move, react to force, and collide dangerously. Odyssey is betting on the same approach: one universal foundation built from video observation, followed by a few dozen hours of machine-specific data.

How the model works

The architecture is an autoregressive diffusion transformer. Training draws on three types of data:

• Internet video annotated with time-stamped events and schema-checked — this provides breadth of understanding • Gameplay recordings synced with keypresses and mouse movements — this provides the link between action and consequence • Rigid-body interaction simulations with captions and metadata — this provides precise physics with known ground-truth outcomes

During training, the model becomes autoregressive through teacher forcing: it's shown real previous frames and forbidden from peeking into the future. Distillation (distribution-matching and adversarial distillation) is then applied, turning the slow diffusion process into a fast one that runs in just a few steps. This lightweight version is what runs in the browser.

There are two models: Odyssey-3 at 832×480 resolution and Odyssey-3 Pro at 1280×720. You walk through the generated world in first- or third-person view, move the camera independently, and inject events during generation to see how the world reacts.

Results: records with caveats

On the Physics-IQ Verified benchmark (from Anates Labs and DeepMind), Odyssey-3 Pro scored 66.1% in video-to-video mode — but that's the best-of-8 setting, where the best result out of eight attempts is selected. Regular scores are more modest: 63.4% with improved prompts, and 52.8–54.7% in image-to-video mode.

On WorldMark (which measures control adherence, scene quality, and world memory), Odyssey took first place in the third-person categories (79.0 and 76.3%), but ranked only third in real-world first-person settings (80.6 versus leader Lyra 2.0 at 84.4). The margin is sometimes as small as 0.2 points. Importantly, the WorldMark evaluations were conducted by the company itself, not by an independent party.

From video to hardware

To adapt the model to a specific robot, the base model stays frozen and a small trainable module is added that translates representations into commands. Here's what's been demonstrated:

• Manipulators: with tens of hours of demonstrations, they perform tasks like "pour cereal into a bowl" and "close a box of screws," and show recovery behavior without explicit training (reorienting the grip after a miss). Partnership with Poke & Wiggle • Humanoids (Flexion): policies generalize better to changes in lighting and environment than VLA models • Cars: after 20 hours of simulated driving, they drive on real Indian streets, predicting waypoints. Policies trained in full simulation cover 77% of the distance achieved by policies trained on real data • Drones: navigation after tens of hours of simulation, stable obstacle avoidance (in a simulated room) • Games: policies play GTA V and transfer to horseback riding in Red Dead Redemption 2 and motorcycles in Sleeping Dogs without additional training • Multi-camera data: after 100 training steps, generates video from three of a car's front-facing cameras

The world model also works as a training ground for other agents. The PROWL project has agents hunt for the model's weak spots, which improves the model, and the more reliable worlds then train agents further. Alongside it are Agora-2 (multiple agents in a single world) and Starchild-1 (extending beyond visual learning).

What should give you pause

The record-setting 66.1 figure is a best-of-8 result, and the WorldMark evaluations were run by the company itself. There's an element of cherry-picking: you see the successful clips, not the failures. Physics-IQ tests how well an experiment continues, not the quality of robot control. Many demos run in simulation, and "tens of hours of data" is still a substantial amount. No public pricing, licensing, or API documentation has been disclosed; developer access is by request only.

In the authors' defense, they're upfront that world models still lag roughly two orders of magnitude behind LLMs, and they acknowledge that robustness across different robots remains an open question.

Odyssey-3 is an attempt to do for the physical world what LLMs did for text: one pretrained foundation, followed by a few hours of task-specific data, and the task is ready. For now, it's a mix of real achievements and promises that still need verifying. The preview can be tried in your browser.