One model, many bodies

On 17 September Odyssey introduced Odyssey-3, a world model built to control robots, humanoids, autonomous vehicles and drones from the same underlying system. In humanoid tests with Flexion the company said tens of hours of teleoperation data were enough to train control policies that opened containers and moved objects, and that they generalised better than the vision-language-action baselines it tested against. It showed the same model on robot arms, in simulated driving, in drone navigation and in video games, and said a public release would follow within weeks.

Every robot maker now runs on somebody else's brain. The layer that records, simulates and grades what the brain does is the one nobody can swap out.

It arrives on top of a year in which the model layer of physical AI compounded faster than the machines. In April, Physical Intelligence published pi-0.7, a seven billion parameter vision-language-action model it describes as steerable and general: it performs tasks it was never trained on by recombining skills, operating a kitchen appliance it had seen twice, folding laundry on a robot for which no folding data existed. In June, NVIDIA released Cosmos 3, a world foundation model that unifies synthetic world generation, vision reasoning and action simulation, in a 16 billion and a 64 billion parameter variant, and made GR00T N1.7 generally available with a new reasoning backbone and pretraining on twenty thousand hours of egocentric human video alongside robot demonstrations, with a second generation previewed for the end of the year. NVIDIA also reports that adding synthetic motion data from its simulation pipeline lifted task success by roughly forty percent over real data alone; that figure is the company's.

Notice what each announcement is actually selling. Not a robot. A model that treats the robot as a peripheral, and a pipeline for the data, real and synthetic, that the model eats. The hardware maker's moat is thinning in real time: if the same brain drives a humanoid, an arm and a car, the differentiated part of a robot company is the body, and bodies are a manufacturing business with manufacturing margins.

The durable part of the stack is the one every program touches and none can replace on a release cycle. The teleoperation and data operations that produce those tens of hours. The simulation and synthetic data that multiply them. The recording, replay and evaluation layer that tells a fleet operator what the model did, why, and whether the next version is safe to ship. Each of these gets more valuable when the model improves, because a better model raises the bar on what has to be measured before it is trusted with a forklift.