OpenAI Robotics Launch: Giving AI a Body, Not Just a Brain

Aditya Ramesh, the man who led development on OpenAI’s video generator Sora, is now the head of their new robotics division. Last week, we covered AgentKit, OpenAI’s move to give AI a mouse and keyboard to manipulate software. This week, they’re giving it a body. OpenAI has officially launched OpenAI Robotics, a full-fledged division dedicated to building physical machines. This isn't an experiment. This is a declaration. The story here starts with a breakup. Until early 2025, OpenAI was partnered with humanoid robotics company Figure.

That partnership ended. At the time, it was corporate boilerplate. Now we know why. Figure’s CEO, Brett Adcock, said general AI models couldn't meet the specific hardware needs of a robot. He believed in vertical integration — building the hardware and software together, from the ground up. OpenAI disagreed. Violently. OpenAI’s entire strategy is the exact opposite. Their philosophy: "Build the brain first, then grow the body." They are betting everything on the idea that a sufficiently powerful, generalized AI model — a foundation model for the physical world — can learn to operate any body.

They don't want to build a better robot. They want to build the single operating system that runs inside ALL robots. Sam Altman said it himself: "We’re no longer content with just building the brains for other companies' hardware." This isn't a product launch. This is a platform play for the entire physical world. The robotics market is a mess of fragmented, custom-built operating systems. It's a fifty billion dollar industry waiting for its "Windows" moment, its "Android" moment.

That's the prize OpenAI is after. The market for the robot OS is growing to nearly a billion dollars this year alone, and analysts agree it’s a winner-takes-most dynamic. OpenAI isn’t just trying to compete with Figure or Boston Dynamics. It’s trying to make them hardware vendors for the OpenAI ecosystem. The first robot they build won't be the main product. It will be the reference device. The first Google Pixel phone for their new Robotics OS. But they've tried this before.

And failed. In 2021, OpenAI shut down its last robotics team. The reason? Not enough data. Robots are expensive. Running them for millions of hours to gather training data is slow and impractical. So what's different now? Two things. First: Simulation. Aditya Ramesh isn't the head of robotics by accident. The same world models that allow Sora to generate impossibly realistic videos can be used to run millions of simulated robot trajectories. They can train the AI brain in a virtual world that is finally good enough to transfer to the real one.

They're hiring simulation realism engineers just for this. Second: Multimodal models. Their new AI can fuse data from cameras, LiDAR, depth sensors, and even tactile sensors. It can see, and now, it can feel. They're solving the data problem by getting smarter about the data they collect, and by generating most of it virtually. The eleven new job roles aren't for research. They're for custom actuator design and data acquisition operations. They are building the machine.

For years, the question in robotics has been about hardware. Who can build the most stable, most agile, most capable machine? That was the wrong question. OpenAI is betting that hardware is about to become a commodity. The real fight isn't about who builds the best robot. It's about who builds the brain that makes the hardware smart. It's about the operating system. Other companies are building better horses. OpenAI just started manufacturing the internal combustion engine.

The battle for robotics isn't about who builds the best hands. It's about who writes the code that moves them. And OpenAI just announced they're not just entering the race… they plan on owning the racetrack.