OpenAI Will Build Its Own Humanoid Robot, Altman Confirms
OpenAI is building a humanoid robot itself.
Chief executive Sam Altman confirmed the plan on the Sources podcast. Asked whether the company would build the hardware, he answered, "We will." Engadget
Altman said the human form is logical for a robot because the world is very much designed for people. Doors, keyboards and tools are designed around the human body.
He said figuring out the brain that makes the robot work is more important than details of what the robot looks like.
On timing, Altman said OpenAI will have a cool humanoid robot demo in 2027. He does not expect robots walking the streets for a few years longer.
The hiring detail describes a full hardware and software program. OpenAI is recruiting for robotics roles spanning motion control, state estimation, actuator control, whole-robot behavior, and machine-learning infrastructure to train physical AI systems.
Its listings describe work on custom electromechanical actuators including motors, transmissions, sensing and thermal systems. They also describe manufacturing and supply chain for hardware at scale.
OpenAI has prior work in robot learning. Its Dactyl research project taught a robotic hand to manipulate real objects with training done entirely in simulation, later applied to a Rubik's Cube.
The confirmation follows earlier reporting. In early September, coverage described OpenAI as building a humanoid robot itself, not just the physical AI to power it Forbes. Altman separately told Bloomberg that the 'humanoid robot moment' hasn't arrived yet but is coming soon Automate.
Separately, OpenAI debuted GPT-6 Astra in early September 2026, described as the company's latest AI model Bloomberg. In January, the company said it was investing in Merge Labs to support new brain-computer interfaces that bridge biological and artificial intelligence OpenAI. In September 2026, Altman told employees the company was open to slowing development of its AI Reuters. On October 15, 2026, he backed a push for AI regulation and said the world was right to be afraid Reuters.
Looking at what this means for practitioners, the split Altman draws between brain and body is the right place to focus. Locomotion and grasping can be shown in controlled conditions. Reliable autonomy is harder. State estimation under contact, whole-robot behavior after a slip or collision, and policies that transfer from simulation to varied objects and lighting remain the limiting factors. A controlled demo in 2027 is plausible. Unattended operation on streets a few years later is consistent with that difficulty.
In my view, the listings deserve close reading. Motion control, state estimation and actuator control plus custom motors, transmissions, sensing and thermal systems plus manufacturing and supply chain is a vertically integrated program. Anyone can render a humanoid. Sustaining consistent hardware at scale is harder. If OpenAI holds that course, it will be tested on factory discipline as much as model training. The long-term possibility is clear enough. A physical platform fitted to human environments could absorb software gains for years.
