Odyssey, a three-year-old AI lab founded by two self-driving car veterans, just released Odyssey-3, a single AI system that can control robot arms, humanoid robots, self-driving cars, and drones, and even play video games like Grand Theft Auto V. The same underlying model powers all of it. That is the real news here, not any one of the demos on its own.
Here is the simple idea behind it. Most robots today are trained one task at a time. You want a robot to pack a box, you show it thousands of examples of boxes being packed. You want it to pour coffee, you start over with thousands of coffee-pouring examples. That is slow and expensive, and it is a big reason robots have stayed stuck in narrow, repetitive jobs.
Odyssey trained its system by having it watch enormous amounts of video of the real world, the same way a toddler learns that dropped objects fall and that people react when you bump into them. Once that general understanding is built in, the company says it only needs a small add-on layer, trained on tens of hours of task-specific footage, to teach the system a new machine or a new job. That is a fraction of what task-specific robot training usually takes.
The company is not doing this alone. Flexion, a Zurich robotics startup that raised 50 million dollars earlier this year, is building humanoid robot control on top of Odyssey-3. Poke and Wiggle, a robot data and benchmarking firm, is testing how well the system holds up across different robot bodies and environments. Both partnerships suggest Odyssey is positioning itself as an infrastructure layer that other robotics companies build on, rather than a robot maker itself.
This announcement lands three months after Odyssey raised 310 million dollars at a 1.45 billion dollar valuation, with Amazon among the backers. Amazon's involvement is not incidental. Odyssey has committed to running its models on Amazon's own AI chips, giving Amazon a stake in an alternative to Nvidia's hardware for robotics.
Odyssey is not the only company chasing this idea. Google DeepMind has its own world model called Genie 3, Nvidia has been pushing a platform called Cosmos for the same purpose, and Fei-Fei Li's World Labs recently raised a billion dollars to pursue nearly identical goals. Whoever builds the most reliable general-purpose model for physical understanding will likely become the default supplier that robot and vehicle makers plug into, the same way cloud computing consolidated around a handful of providers.
A word of caution: everything above comes from Odyssey's own announcement, with no independent testing of the numbers yet, and the model itself is not publicly available. It is still a set of research demos and early partnerships, not a product you can buy today.
Still, if even a version of this works at scale, the cost of teaching a machine a new physical task drops sharply. That matters for any business that has looked at warehouse robots, delivery drones, or automated inspection and decided the setup cost was not worth it. The barrier has always been the mountain of task-specific training data required. If that mountain shrinks to a few days of footage, the economics of automation change for companies far smaller than Amazon or Tesla.