Odyssey Unveils Odyssey-3 World Model to Advance Physical AI

Artificial intelligence startup Odyssey has unveiled Odyssey-3, its latest foundation world model, as the company expands its efforts to develop AI systems that understand and predict how the physical world changes over time. Announced on October 8, 2026, the model is designed to learn from visual observations and predict how objects move, interact and respond to different conditions. Unlike conventional AI systems that primarily generate text, Odyssey-3 focuses on modelling physical environments and the relationships between objects and actions. The company believes this technology could help AI systems develop a more practical understanding of their surroundings, enabling them to perform tasks in physical and virtual environments. Odyssey has made a research preview available for developers interested in exploring its capabilities.

At the centre of Odyssey-3 is a technology known as a world model. This type of AI system learns patterns about how an environment behaves and uses that information to predict what may happen next. For example, a world model could estimate how a moving object will respond when pushed, how a vehicle might behave during a turn or how changes in an environment could affect a robot’s next action. Odyssey describes its model as an autoregressive diffusion transformer that learns physical dynamics and cause-and-effect relationships from visual data. By predicting how situations evolve, the system can help developers simulate possible outcomes before applying a decision in the real world. This approach could become particularly useful in robotics and autonomous systems, where understanding the consequences of an action is essential.

Odyssey says the new model represents a significant improvement over its previous systems. According to the company’s announcement, Odyssey-3 Pro achieved state-of-the-art performance on the Physics-IQ Verified benchmark and ranked first in three of the four WorldMark categories in the company’s evaluations. These results are intended to demonstrate the model’s ability to represent physical interactions and generate predictions that align with real-world dynamics. However, the reported results come from Odyssey’s own evaluations, and independent testing will be important to determine how well the model performs across different environments and practical applications. Benchmark performance alone does not guarantee that an AI system will operate reliably when faced with unpredictable real-world conditions.

The potential applications extend across several industries. In robotics, developers could use Odyssey-3 to generate simulated environments in which machines learn how to manipulate objects and respond to changing surroundings. Autonomous-driving researchers could investigate driving situations, while drone developers could explore navigation and movement in complex environments. The model could also support AI training for video games and virtual simulations, where realistic interactions can make digital environments more responsive. One major attraction of world models is their potential to reduce dependence on expensive physical testing by allowing developers to investigate more situations in simulation. Nevertheless, simulated environments must accurately reflect real-world physics, and systems trained in simulation still require careful testing before they can be trusted to control physical machines.

Odyssey’s launch comes as technology companies increasingly invest in physical AI, an area focused on enabling intelligent systems to understand and interact with the physical world. The company was founded in 2023 around the idea that general-purpose world models could provide a shared foundation for different machines and tasks, rather than requiring separate AI systems for every application. In June 2026, Odyssey announced a $310 million Series B funding round at a reported valuation of $1.45 billion, with investors including Amazon, AMD Ventures, Google Ventures, EQT and In-Q-Tel. The introduction of Odyssey-3 advances the company’s effort to turn world modelling into a practical development platform. Its long-term success will depend on the accuracy of its predictions, the resources needed to run the system and whether developers can use it to build reliable AI applications outside controlled simulations.