Odyssey-3: A Generative World Model for Interactive Environments

The California-based AI company Odyssey has made its world model, Odyssey-3, publicly available in a research preview. This generative model can create interactive environments from text prompts in real-time, allowing users to explore and interact with them from different perspectives.

Odyssey-3 is designed not only to generate virtual worlds but also to serve as a foundation for controlling various systems, including robotic arms, drones, and characters in video games. The company aims to use the same world model across different tasks, streamlining development and reducing complexity.

The public research preview of Odyssey-3 offers interactive environment generation capabilities, with users able to generate environments from text descriptions and explore them in real-time. This includes choosing between first-person and third-person perspectives, moving through the generated world, triggering events, and watching the model respond in real-time.

Odyssey-3 is built on an autoregressive diffusion transformer that continuously generates new video frames based on previous frames and user actions. During training, the model learns physical relationships and cause-and-effect from visual observations, allowing it to simulate complex environments with realistic physics.

The base Odyssey-3 model has 14 billion parameters and generates video at a resolution of 832 × 480 pixels, according to official benchmark submissions. The more powerful Odyssey-3 Pro supports higher resolutions up to 1280 × 720 pixels.

Odyssey-3’s performance on physics benchmarks is impressive, with the company claiming top scores in several categories. However, there are caveats to these claims: the results come from a single test run where a selection method picked one of eight generated videos for each task, and do not meet the benchmark rules’ requirement for four test runs.

The WorldMark benchmark shows Odyssey-3 ranking first in three out of four categories based on its own evaluation. In First-Person Stylized, it achieved 77.2 points; in Third-Person Real, 79.0 points; and in Third-Person Stylized, 76.3 points. However, in First-Person Real, it placed third with 80.6 points.

Odyssey-3’s capabilities extend beyond generating virtual worlds to controlling robotic arms, drones, and characters in video games. The company has demonstrated an AI built on Odyssey-3 controlling multiple robotic arms, recovering from failed grasps without additional training data.

The same world model is being used for different applications, including robotics, autonomous driving, and video games. For each application, the model gets paired with a specialized controller that translates its predictions into concrete commands. This approach aims to reduce development time and increase efficiency across various tasks.

Odyssey has partnered with Swiss robotics company Flexion to develop humanoid robot controllers using Odyssey-3. These controllers are said to perform more reliably than comparison models, even in changing conditions. The company also demonstrated a drone controller trained on simulated flight data, capable of dodging obstacles and flying to specific targets.

The public research preview offers interactive environment generation capabilities for now, with further work needed for robotics and autonomous systems. Odyssey’s goal is to use its world model as a foundation for controlling various systems, reducing complexity and increasing efficiency across different tasks.

Odyssey was founded in 2023 by Oliver Cameron and Jeff Hawke, who have since raised $310 million from investors including Amazon and AMD Ventures. The company aims to revolutionize the development of AI systems with its generative world model, making it easier for developers to create interactive environments and control various systems.