California-based AI company Odyssey has released a free public research preview of Odyssey-3, a generative world model that creates interactive environments in real time.

California-based AI company Odyssey has launched a public research preview of its generative world model, Odyssey-3. Users can generate interactive environments from text prompts, explore them in real time, and test the model for free using an online demo running on Odyssey-3 Flash. Founders Oliver Cameron and Jeff Hawke originally unveiled the model on September 15, with applications focused on robotics, autonomous driving, and video games.

Model Architecture and Training Data

Odyssey-3 is built on an autoregressive diffusion transformer that continuously generates new video frames based on previous frames and user actions. The base model features 14 billion parameters and generates video at an 832 by 480 pixel resolution, while Odyssey-3 Pro supports 1280 by 720 pixels. Training data included internet videos with event descriptions, video game footage paired with keyboard and mouse inputs, and simulated physical interactions. An extra training technique reduces the number of required compute steps to allow real-time generation.

Benchmark Scores and Testing Caveats

According to the company, Odyssey-3 Pro scores 66.1 points on the video-to-video benchmark from Physics-IQ Verified, a test evaluating physical behavior across fluid mechanics, optics, solid mechanics, magnetism, and thermodynamics. However, that top score comes from a single test run using a selection method that picked one of eight generated videos for each task, falling short of benchmark rules requiring four test runs with standard deviation reported. Without the selection method, Odyssey-3 Pro averaged 63.37 points across four runs. On the WorldMark benchmark, Odyssey-3 ranked first in three of four categories based on its own evaluation: First-Person Stylized at 77.2, Third-Person Real at 79.0, and Third-Person Stylized at 76.3, while placing third in First-Person Real with 80.6 points.

Applications in Robotics and Gaming

Odyssey intends to use the world model across multiple tasks by pairing it with specialized controllers. In company tests, an AI built on Odyssey-3 controlled multiple robotic arms after training on a few dozen hours of demonstration data, and worked with Swiss robotics company Flexion on humanoid robot controllers. Odyssey also developed a drone controller trained on simulated flight data and demonstrated an AI autonomously playing Grand Theft Auto V, with a controller trained on roughly two hours of GTA footage transferring skills to Red Dead Redemption 2 without extra training.