Odyssey opens research preview of interactive world model Odyssey-3

Public preview and model design

California based AI company Odyssey has launched a public research preview of its world model Odyssey-3. Users can generate interactive environments from text prompts and explore them in real time from first person or third person views, move through the world, trigger events, and watch the model respond. The free demo runs on Odyssey-3 Flash, and developers can apply for API access.

Odyssey says Odyssey-3 is built on an autoregressive diffusion transformer that generates new video frames from previous frames and user actions. 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 compute steps for real time generation. The base model has 14 billion parameters and generates 832 by 480 video, while Odyssey-3 Pro supports 1280 by 720 pixels.

Benchmark claims include caveats

On Physics-IQ Verified, Odyssey-3 Pro scores 66.1 points according to Odyssey. The video to video benchmark covers fluid mechanics, optics, solid mechanics, magnetism, and thermodynamics by asking models to continue videos of real world experiments and comparing outputs with actual outcomes.

Odyssey says that score came from one test run with a selection method that picked one of eight generated videos per task. Benchmark rules require four runs and a standard deviation for record claims, and the reported result does not meet that bar. Without selection, Odyssey-3 Pro averaged 63.37 points across four runs. Both results appear on the leaderboard submitted by Odyssey. On WorldMark, Odyssey says it ranks first in First-Person Stylized at 77.2, Third-Person Real at 79.0, and Third-Person Stylized at 76.3, and third in First-Person Real at 80.6.

From robots to game characters

Odyssey wants one world model for robotic arms, drones in simulation, and video game characters, paired with controllers that turn predictions into commands. In company tests, an AI built on Odyssey-3 controlled multiple robotic arms after a few dozen hours of demonstration data and recovered from failed grasps that were not in training data. For humanoid robots, Odyssey is working with Swiss robotics company Flexion.

Odyssey also built a drone controller trained on simulated flight data, saying the drone dodged obstacles and reached targets in a virtual indoor environment. It showed an AI playing GTA V autonomously, steering vehicles and fighting enemies. A controller trained on about two hours of GTA footage transferred to Red Dead Redemption 2 without extra training. Odyssey also plans to use the model for AI agents that complete natural language tasks through their own actions.

Sources: The Decoder