Odyssey - 3

Odyssey is a research lab working on world models: AI systems that build an internal understanding of how reality behaves. Rather than treating video, physics and agent behaviour as separate problems, a world model learns the properties and constraints of objects (physics), how objects and systems move and change (dynamics), how actions influence what happens next (cause and effect) and how agents pursue goals and interact with each other. Odyssey argues that this shared understanding is what lets robots perform useful work, lets AIs learn through experience, and lets humans explore and test ideas inside simulated worlds.

Odyssey-3 is the company's most advanced foundation world model, described as materially advancing the state of the art in physical accuracy. Beyond Odyssey-3, the lab has published Agora-2, a multi-agent world model that lets multiple human or AI participants share and interact within the same real-time simulation, and PROWL-2, a reinforcement learning framework in which a team of agents and its world model each learn from their own curriculum in a single training loop. Starchild-1 extends the line toward systems that learn from richer multimodal interaction rather than visual observation alone.

The intended applications span robotics, humanoids, self-driving, drones, defense, gaming, energy, cyber and general AI training, and the team draws on experience from DeepMind, Waymo, Tesla, Meta, Apple and xAI. Odyssey has raised a $310M Series B. Product access is offered through a Try Odyssey-3 demo and a separate API access request rather than published self-serve plans, so commercial terms are arranged directly with the company.