Odyssey pitches Odyssey-3 as one world model for robot arms, humanoids, cars and drones, trained on tens of hours per machine

AI lab Odyssey said in a blog post by co-founders Oliver Cameron and Jeff Hawke that Odyssey-3, an autoregressive diffusion transformer it calls its most powerful foundation world model, can be adapted to control robot arms, humanoids, cars and drones, and to generate training environments and play video games.

The company said a learned action decoder attached to the model translates its internal representations into machine commands, that tens of hours of demonstrations were enough to teach robot arms several manipulation tasks with recovery behaviours absent from the training data, and that humanoid partner Flexion built real-time control policies on top of it from tens of hours of teleoperation data.

Robot Newspaper reported that Odyssey also claims a score of 66.1 on the Physics-IQ Verified benchmark; that figure was not on the post MEKAWIRE opened and comes from the company's own evaluation. In MEKAWIRE's coverage, Odyssey-3 lands in the same month as Google DeepMind's Gemini 4 Argon for Robotics and NEURA's Neuron AI, a sign that world-model vendors now compete directly with robot makers for the robot's 'brain'.