California-based AI startup Odyssey has launched a public research preview of Odyssey-3, a generative world model capable of creating interactive 3D environments from text prompts in real time. Built on an autoregressive diffusion transformer, the 14-billion parameter base model simulates physical processes and predicts environmental changes based on user inputs. The release includes a free online demo running on Odyssey-3 Flash alongside API access for developers.

In benchmark evaluations reported by the company, Odyssey-3 Pro achieved top results in physical simulation tests, scoring 66.1 points on the Physics-IQ Verified video-to-video benchmark, though without custom selection methods it averaged 63.37 points across standard runs. The model also ranked first in three of four categories on the WorldMark benchmark, which measures control instruction compliance, visual quality, and temporal consistency.

Targeting applications across robotics, autonomous driving, and video games, Odyssey pairs the world model with specialized controllers to translate predictions into physical commands. According to the company, an Odyssey-3 system successfully controlled multiple robotic arms after training on just a few dozen hours of demonstration data.

Why it matters

  • Real-time generative world models provide a scalable simulation base for training autonomous systems and robotics with minimal real-world data.

  • Interactive physics simulation models enable new possibilities for real-time environment generation in video games and virtual testing.

Source: the-decoder.com