New World Model Odyssey-3 Scores 66.1 on Physics Benchmark, Surpassing FLUX 3, Opens Free Trial

Beating AI News Flash: World model company Odyssey officially releases Odyssey-3 and opens free trial play for the Flash version. It can generate explorable worlds in real time from text. As users walk and turn the camera, the visuals change accordingly. The team also demonstrated experiments using the same model for robotic arms, humanoid robots, cars, and game characters.

Odyssey-3 first learns object motion and the consequences of actions from internet videos, gameplay recordings, and physics simulations. When used for robots or cars, the parameters of the pretrained main model remain unchanged. The team only needs to additionally train small control modules to translate the knowledge learned by the model into operational commands. Different machines still require separate adaptation, but there is no need to retrain the entire model.

Odyssey used only 20 hours of simulated driving data to train a dedicated control module, enabling a car to drive on real roads in India. In testing, the simulated-data training approach achieved a distance between two safety driver takeovers that was about 77% of that of the real driving footage training approach.

A humanoid robot developed in collaboration with Flexion, after training with dozens of hours of teleoperation demonstration data, was able to continue performing tasks under changing lighting conditions. In game experiments, the control module learned from about 2 hours of GTA V gameplay footage. After that, with no additional training, it was able to make a character ride a horse and move in Red Dead Redemption 2.

On the Physics-IQ Verified physical video prediction test, Odyssey-3 Pro scored 66.1 points, surpassing FLUX 3 [large]'s 64.35 points, both based on best-of-8 results. Odyssey also announced the WorldMark evaluation results, with the model taking first place in three of four interactive scenario categories.