Odyssey Launches Odyssey-3 for Physical AI

Odyssey-3 AI Humanoid Robot Manipulation

Odyssey has launched Odyssey-3, its most powerful foundation world model, to support AI systems operating across physical and virtual environments. The company announced the model on October 8, 2026, marking a new step toward general-purpose physical intelligence.

Unlike conventional AI models built for specific tasks, Odyssey-3 learns how environments behave and how actions change outcomes. Consequently, developers can use its learned understanding of physics, movement, and cause-and-effect across different machines and applications.

The model targets a broad range of uses, including robotics, autonomous driving, drones, humanoid systems, AI training, and interactive gaming. Moreover, Odyssey aims to help developers adapt a single underlying model to different tasks using less task-specific training data.

Odyssey-3 Targets Robotics, Vehicles and Drones

Odyssey-3 uses an autoregressive diffusion transformer to predict how objects move and environments change over time. As a result, the model can generate interactive simulations that respond to user actions and changing conditions.

The company has demonstrated the model across six areas. These include robotic arms, humanoid robots, vehicles, drones, AI agent training, and video games. In addition, Odyssey says the model can generate training environments where AI agents learn through actions and their consequences.

Apple and LG Reportedly Partner on New Smart Home Devices

One notable demonstration involved autonomous driving on Indian roads. Odyssey trained a driving policy using approximately 20 hours of simulated driving data while keeping the foundation model frozen. However, this result describes the company’s demonstration and does not establish readiness for unrestricted public-road deployment.

Odyssey has also collaborated with Flexion on humanoid robotics. The resulting control policies perform tasks in real time and, according to Odyssey’s evaluations, handle certain environmental changes better than the tested baseline systems. Meanwhile, robot-arm demonstrations showed recovery behaviours, including responses to unsuccessful grasps.

Benchmark Results and Research Access

Odyssey reports improvements in physical accuracy and interactive world generation. Its Odyssey-3 Pro version scored 66.1 on Physics-IQ Verified’s video-to-video benchmark, which evaluates predictions involving physical phenomena. These include fluid dynamics, optics, solid mechanics, magnetism, and thermodynamics.

Furthermore, Odyssey reports that the model ranked first in three of WorldMark’s four evaluated categories. Those categories assess generated environments for visual quality, control-following, and world memory. However, benchmark performance alone does not guarantee reliable operation across all real-world tasks.

The company has released a research preview of Odyssey-3 for users to explore interactive environment generation. Physical AI developers can also contact Odyssey about applying the model to robotics, autonomous vehicles, drones, and other systems.

Ultimately, Odyssey-3 reflects a broader effort to build AI that learns reusable representations of how the world works. If these capabilities transfer reliably across machines and environments, they could reduce the amount of task-specific training needed for physical AI. Nevertheless, developers will need to conduct further testing to establish safety, reliability, and performance beyond controlled demonstrations.