OpenAI is showing renewed interest in robotics, with recent hires and job listings pointing toward work on humanoid systems and AI models that can learn from physical interaction. The push suggests the company is exploring whether progress toward AGI depends on systems that can understand and act in real-world settings.
Robotics is back on OpenAI’s agenda
The clearest signals come from recruitment. OpenAI has recently hired researchers with experience building AI algorithms for humanoid and other robots, according to the source article. People familiar with the company’s efforts also say OpenAI is recruiting specifically for humanoid robots, including machines with partial or full human form.
One source who works in cutting-edge robotics says the company has started training AI algorithms that can better interpret the physical world. The goal, as described in the source, is to support robots that can navigate and perform tasks.
Recent hires help show the direction of travel. Chengshu Li joined OpenAI in June 2025 from Stanford University, where he worked on robotics projects including a benchmark for measuring humanoid robots able to perform many household chores. His dissertation focuses on benchmarks and on robots with a partly humanoid body: two arms, but wheels instead of legs.
The source also notes that two other researchers from another robotics lab have joined OpenAI, according to their LinkedIn profiles. A professor at a third lab that works on humanoids says one of their students was recently recruited as well.
What the job listings reveal
OpenAI declined to comment on its recruiting or robotics plans. Still, its job listings describe a robotics team interested in systems that can be trained through teleoperation and simulation.
One opening requires expertise in teleoperation and simulation. In this context, teleoperation means a human operator controls a robot’s limbs while doing chores, giving an algorithm examples it can learn to imitate. Simulation lets an algorithm learn in a virtual physical environment before a robot is expected to perform in the real world.
The listing also mentions Nvidia Isaac, a simulation tool widely used for training humanoids. That detail matters because humanoid robots need more than language ability. They need control systems that can connect perception, movement and physical manipulation.
Another job posted in the past few weeks calls for a mechanical engineer with experience prototyping and building robot systems using touch and motion sensors. A roboticist cited in the source says this could point either to OpenAI building its own robot or to the company developing teleoperation systems for robot training.
The same job calls for “experience designing mechanical systems intended for high volume (1M+), problem-solving on assembly lines,” which suggests interest in systems that could be mass-produced or used in manufacturing.
The AGI link is explicit
The robotics listings all describe the team as focused on “unlocking general-purpose robotics and pushing towards AGI-level intelligence in dynamic, real-world settings.” That wording connects robotics directly to OpenAI’s broader ambition around artificial general intelligence.
The source defines AGI as AI that exceeds human intelligence. A renewed robotics effort suggests OpenAI may believe that reaching that goal requires algorithms that can engage with the physical world, not only produce text, code, images or video.
That would mark a return to a field OpenAI had stepped away from. The company did notable robotics research in its early years, including an algorithm that solved a Rubik's cube using a humanlike hand in 2019. It shuttered its robotics effort in 2021 to focus on algorithms including large language models, which later helped drive breakthroughs such as ChatGPT.
OpenAI restarted work on robots last year. The Information reported in December 2024 that the company was considering developing its own humanoid robots.
Why humanoids remain hard
Humanoid robots are drawing more attention because the hardware and software needed to build prototypes has become more common. The source notes that new motors and other components have made working systems cheaper and easier to assemble, while tools such as Nvidia’s Isaac platform have simplified parts of robot control and training.
Investment has also accelerated. Venture capitalists have invested more than $5 billion in humanoid startups since the start of 2024. Morgan Stanley reckons the humanoid industry could be worth $5 trillion by 2050.
But the technical challenge remains large. Humanoids can already perform attention-grabbing demonstrations like dancing, yet they still lack the intelligence needed for complex and unpredictable, or “unstructured,” environments.
To work in those environments, robots need algorithms that go beyond a large language model’s grasp of the physical world. They must control limbs and grippers, walk, handle objects and respond to changing surroundings.
Stefanie Tellex, a roboticist at Brown University, describes the challenge as building AI models capable of “processing high-frame-rate, high-dimensional perceptual input, and producing high-frame-rate, high-dimensional physical outputs.” In plain terms, the model must see detailed, fast-changing information and turn it into detailed, fast physical action. Tellex is not familiar with OpenAI’s specific plans.
A crowded race for physical AI
OpenAI already has leading models for conversation, reasoning, coding, and image and video generation, according to the source. In robotics, though, it will face serious competition.
Humanoid startups including Figure, Agility and Apptronik have emerged over the past few years. Major AI companies including Tesla and Google are also investing in developing and testing humanoids.
Tellex is skeptical that OpenAI automatically enters the field with a decisive edge. “I don’t see them having any magical advantage over anyone else,” she says.
The broader timing is important. The source says disappointment around OpenAI’s GPT-5 is part of a wider realization that reaching humanlike intelligence may require new research directions. Tellex puts the shift bluntly: “They've asymptoted on GPT-5,” she says. “They need to move towards the physical world.”
That is the central significance of OpenAI’s robotics push. The company’s next frontier may not be only a larger model or a stronger chatbot. It may be an AI system that can learn from people, test actions in simulation, control machines and operate amid the uncertainty of the real world.