Odyssey-3 extends one world model across robots, cars and drones

Odyssey has introduced Odyssey-3, a foundation world model designed to provide a shared physical-world representation across different machines and control systems.

Rather than training a separate base model for each embodiment, Odyssey keeps the pretrained world model frozen and adds smaller action decoders or control policies that translate its internal representations into commands for a specific robot, vehicle or drone. The company says Odyssey-3 was trained on broad visual observations of the world to learn concepts including physics, dynamics, cause and effect and human behaviour.

For robot arms, Odyssey says the model can learn control from tens of hours of demonstrations, including recovery behaviours that were not present in the demonstrations themselves. The company is working with Poke & Wiggle to benchmark how well that transfer holds across different robot bodies, viewpoints and control schemes.

Odyssey-3 is also being used as a base model for Flexion’s humanoid autonomy stack. Flexion adds its own robot-learning and whole-body-control systems, with the resulting policies trained from tens of hours of humanoid teleoperation data. Odyssey says those policies handled environmental changes such as lighting better than the VLA baselines tested in its evaluations.

Odyssey trained a small policy on 20 hours of simulated driving data while keeping the world model frozen, then ran it closed-loop on public roads in India. The company says the simulation-trained policy travelled around 77% as far between safety-driver interventions as comparable policies trained on real driving footage.

The same approach has also been applied to drone navigation and game-playing agents. The drone work has so far been demonstrated in simulation, while game policies trained in GTA V showed some transfer into Red Dead Redemption 2 and Sleeping Dogs without additional training on those titles.

Odyssey-3 acts as the shared pretrained foundation, while each system still needs a control layer trained around its own actions and dynamics.

Odyssey was founded in 2023 by Oliver Cameron and Jeff Hawke. In June, the company announced a $310 million Series B at a $1.45 billion valuation, led by Natural Capital with participation from Amazon, GV, AMD Ventures, EQT, IQT and others.

Have a robotics update Korthos should review? Send news, deployments, product releases, funding rounds, research, or media to [email protected] or reach out on X at @agkorthos.