Odyssey-3 is the latest foundational world model released by Odyssey, which it officially calls its most powerful world model to date. Its core purpose is to learn physics, dynamics, and causality, thereby predicting how objects move and interact in space and how situations evolve over time, and it can be used by developers to simulate environments and train policies for different physical systems.
In terms of main capabilities, Odyssey-3 is a learned dynamical system implemented as an autoregressive diffusion Transformer. It can generate embodied environments based on prompts and predict in real time how the environment changes when a person or agent takes actions or introduces events; it uses prior observations and the latest input to make predictions. The current research preview supports first-person and third-person navigation as well as independent camera movement.
In terms of evaluation performance, Odyssey officially says that Odyssey-3 Pro scored 66.1 on the Physics-IQ Verified video-to-video benchmark, the highest reported score; its image-to-video score is 54.7. Physics-IQ tests cover physical behaviors including fluid dynamics, optics, solid mechanics, magnetism, and thermodynamics. In the WorldMark evaluation, using the benchmark's own description text and the mean of 13 metrics, Odyssey-3 ranks first in first-person stylized, third-person realistic, and third-person stylized environments. WorldMark measures control following, visual quality, and world memory.
In terms of adapting to different physical systems, Odyssey-3's world knowledge can be transferred to a specific machine by training an action decoder or policy on paired observation and action data. Examples officially given include: manipulation tasks can be completed with only tens of hours of robot demonstrations, and recovery behaviors not present in the demonstrations emerge, such as readjusting the gripper orientation after a failed grasp and retrieving an object that fell in an abnormal position; Flexion built a humanoid robot control policy based on Odyssey-3, which outperformed the tested VLA baselines under environmental changes and was still able to perform tasks under lighting changes that caused the baselines to fail; on real roads in India, a driving policy was trained with only 20 hours of driving data while keeping the Odyssey-3 backbone frozen, and the policy used the model's visual representations to predict future waypoints, achieving closed-loop driving; in early experiments with the first three cameras of an autonomous driving dataset, after only 100 training steps, an Odyssey-3 training checkpoint could simultaneously generate driving sequences from three camera viewpoints. In addition, world models can provide agents with a decision-making environment for studying agents' responses to condition changes and task completion; in official demonstrations, agents receive naturallanguage goals and execute them in the world generated by Odyssey-3.
In terms of training and construction, the official introduction states that its training data combines internet videos with temporally localized, pattern-validated event annotations, gameplay recordings with time-aligned keyboard and mouse input, and simulated rigid-body interactions with descriptions and metadata. The model is designed to support dynamic, open-ended interaction; it uses a multi-step video diffusion Transformer and employs temporally resolved prompts and control to guide the unfolding of the world; it is trained via teacher forcing and causal masking for autoregressive extension so that it continues from prior observations and predicts future states based on action inputs; finally, a post-training pipeline of distribution matching and adversarial distillation is used to produce a few-step distilled variant capable of real-time interaction.
In terms of usage and applicable scenarios, the official release provides a research preview in which one can prompt an environment, act within it, and observe the world model's responses; Odyssey-3 Flash is also available to try. If you are working on the development of robots, humanoid robots, autonomous vehicles, drones, or other autonomous systems, and want to explore how a foundation world model can accelerate your work, you can contact the official team.
In terms of limitations, the official release notes that the current preview provides interaction through first-person and third-person navigation and independent camera movement; WorldMark's evaluation measures specific properties of generated worlds, and applying the model to a physical system still requires evaluating the relevant behavior for that machine and its task. Information on fees, pricing, platform support, and specific usage steps is not yet provided in the available materials.
