OpenAI says its humanoid robot needs a brain before a body
Sam Altman says OpenAI will build a humanoid robot, but the company still has no product, price or release date.
OpenAI is working towards a humanoid robot, CEO Sam Altman has said, but he believes the difficult part will be building the machine's intelligence rather than giving it a recognisable human shape. The company has not announced a product, price or release date. 1
OpenAI's robot plan is still early
Speaking on the Sources podcast, Altman said the human form makes practical sense because the world is designed around people. Doors, keyboards, tools and workspaces all assume a roughly human body, so a robot intended to operate in those spaces may benefit from matching that form factor. When asked whether OpenAI would actually build one, Altman answered: "We will." 1
That is a commitment to a project, not a launch announcement. Altman has separately suggested that OpenAI could have a demonstration in 2027, while saying robots walking through everyday public spaces are still several years away. OpenAI has not disclosed what its own machine would look like or how much it would cost. 1
The immediate challenge is the software needed to make a physical machine useful. A capable humanoid must interpret camera and sensor data, understand an instruction, plan a movement, control motors and adjust when the real world refuses to behave as expected. Picking up a mug is a simple example: the robot has to identify it, find the handle, judge the grip and apply enough force without crushing it or dropping it.
OpenAI's current robotics recruitment suggests it is working on both sides of that problem. Its listed roles cover motion control, state estimation, actuator control, whole-robot behaviour and the machine-learning infrastructure needed to train physical systems. The company also appears to be developing custom electromechanical actuators, manufacturing processes and supply-chain capability. 1
The body is not an afterthought
Altman's emphasis on the robot's "brain" should not be mistaken for a claim that the hardware is easy. OpenAI's earlier Dactyl work trained a robotic hand to manipulate real objects in simulation, including a Rubik's Cube. That history shows the same gap that faces every robotics company: software can learn what should happen, but motors, friction and unpredictable objects decide what actually happens. 1
Tesla and XPENG are already pursuing humanoid machines, so OpenAI would be entering a crowded field rather than inventing the category. Its reported near-term goal is more practical than a robotic housemate: general-purpose machines that could help skilled workers build infrastructure. A personal robot for a broader range of household tasks is described as a longer-term ambition. 1
This image is illustrative and shows a humanoid robot from Pexels; it is not an OpenAI prototype. 3
Our opinion
The sensible reading of Altman's comments is that OpenAI is treating robotics as an intelligence problem with a long hardware programme attached. That could be a useful distinction: a convincing demo will need more than a robot that walks across a stage. It will need to understand messy environments, recover from mistakes and perform useful work safely.
For now, the important news is that OpenAI has moved humanoid robotics from speculation into an explicitly stated project. The less exciting but more honest part is that the company still has no public product to show, and the timeline remains measured in years rather than months.