OpenAI's Real Robot Challenge Isn't Hardware—It's the Brain

Sam Altman has a straightforward answer when asked whether OpenAI will build a humanoid robot: yes. But the company's CEO is far less interested in debating what that robot will look like than in solving what he sees as the genuinely hard problem—the intelligence that would animate it.

In an appearance on the Sources podcast, Altman laid out the reasoning for the humanoid form itself. The world, he observed, was built for human bodies. Doors open at a certain height. Keyboards sit at a certain angle. Tools fit certain hands. A robot that needs to move through that world and use those tools should probably match the shape of the thing the world was designed around. The form factor, as he put it, makes sense. But once you've settled on two arms and two legs, you've solved maybe ten percent of the problem.

The real work is teaching a machine to see what its cameras show, understand what it's being asked to do, figure out how to do it, coordinate all the moving parts, and then adjust when reality doesn't cooperate. Pick up a mug. The robot has to identify it, locate the handle, estimate how hard to grip it, apply exactly the right amount of force. Too little and the mug stays on the table. Too much and you're mopping coffee off the floor. That gap between what should happen in theory and what actually happens in the physical world is where most robotics companies get stuck.

OpenAI has been here before. Years ago, the company's Dactyl project trained a robotic hand to manipulate objects—all of it learned in simulation, then transferred to real hardware. Later they applied the same approach to a Rubik's Cube. Both projects revealed the same stubborn truth: software can learn what ought to happen, but physics gets the final word. The company is now recruiting people to work through those exact problems. Its job listings span motion control, state estimation, actuator control, and the machine-learning infrastructure needed to train physical systems. There are currently 27 open robotics positions.

Altman's public emphasis on the brain shouldn't be read as OpenAI treating the hardware as secondary. The company is developing custom electromechanical actuators—motors, transmissions, sensors, thermal systems. Other positions cover manufacturing and supply chain work, the unglamorous infrastructure needed to eventually produce robots at scale. That's substantial investment in what Altman called the less important half.

OpenAI is not alone in this space. Tesla has been developing Optimus for years. XPENG recently demonstrated its IRON humanoid walking off the production line under its own power after assembly. Both companies are proof that even sophisticated software has to wrestle with the complexity of coordinating motors and joints into coherent movement.

The company's near-term ambition is more modest than a robot in every home. Internal job postings indicate OpenAI's short-term goal is building general-purpose robots that can assist skilled workers in infrastructure projects. The longer-term vision involves personal robots capable of handling a wider range of household tasks. Altman told Fortune that a demo is likely in 2027, though he doesn't expect robots walking city streets for several years after that. OpenAI has released no details about what that demo will show, what the robot will look like, or when anyone outside the company might see one.

The company also hasn't announced pricing, a subscription model, a release date, or how it will handle privacy concerns around a robot equipped with cameras and microphones moving through someone's home. Those are the questions that separate a research project from a product. For now, OpenAI is hiring, building actuators, and working on the part Altman says actually matters—the part that decides what the robot does next.