Breaking

Odyssey Previews Odyssey-3 World Model for Robots, Vehicles, Drones and Games

September 15, 2026 at 13:32 EDT

Foundation Models
AI Agents
Robotics

Odyssey has previewed Odyssey-3, a foundation world model designed to turn learned physical understanding into control policies for real and simulated systems. The company says the model can support robot arms, humanoids, cars, drones, AI-training environments and video game agents, with public release planned in the coming weeks.

September 15, 2026 | Odyssey

One world model, many bodies

Odyssey-3 is presented as a foundation world model for controlling robots, humanoids, cars, drones, AI-training worlds and game agents.

target domains named

20h

simulated driving data

77%

sim-only distance vs real footage

$310M

Series B announced in June

Simulation-trained driving, measured against real footage

Distance between safety-driver interventions on real roads, with the real-footage policy set to 100%.

77%Sim-only policy

100%Real-footage policy

The result suggests meaningful transfer from simulated data, while still trailing policies trained on real driving footage.

How the control stack is described

Pretrain

Learn from visual observations.

Decode

Map representations to actions.

Control

Run policies across systems.

Why it matters

The same frozen backbone is shown supporting policies for robots, humanoids, vehicles, drones and games.

What remains open

Full success-rate tables, model size, pricing, API terms and independent Odyssey-3 benchmarks were not provided.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.