OpenAI’s revitalized focus on robotics indicates that the organization perceives the path to artificial general intelligence (AGI)—AI that exceeds human intellect—as potentially requiring the creation of algorithms that can engage with the physical realm.
During its foundational years, OpenAI undertook crucial robotics investigations, introducing an algorithm in 2019 that was capable of solving a Rubik’s cube with a human-like hand. Nonetheless, the firm paused its robotics initiatives in 2021 to concentrate on algorithms such as large language models, resulting in advancements like ChatGPT. OpenAI recommenced its robotics endeavors last year, and by December 2024, The Information disclosed that the organization was contemplating the development of its own humanoid robots.
Stefanie Tellex, a roboticist from Brown University, proposes that creating more efficient robots will necessitate the design and training of AI models that can handle high-frame-rate, high-dimensional perceptual inputs, while generating high-frame-rate, high-dimensional physical outputs—essentially, models that possess the ability to see and act with high precision. Tellex is not specifically knowledgeable about OpenAI’s intentions.
While OpenAI has top-performing models for dialogue, reasoning, coding, and visual content generation, it faces significant competition in crafting algorithms for advanced humanoid robots. Numerous humanoid startups, including Figure, Agility, and Apptronik, have surfaced, and major AI firms like Tesla and Google are also pouring resources into humanoid development and experimentation. “I don’t perceive them having any unique advantage over others,” remarks Tellex.
Humanoids are becoming increasingly popular as the necessary hardware and software for constructing functional prototypes are more readily available. Although still expensive and demanding to create, innovative motors and components have reduced costs and facilitated the building of operational systems. Software platforms such as Nvidia’s Isaac robot development suite have made it easier to write the code essential for controlling and training humanoid systems.
The excitement surrounding humanoids is on the rise. Since early 2024, venture capitalists have invested over $5 billion into humanoid startups. Morgan Stanley predicts that the humanoid sector could reach a valuation of $5 trillion by 2050.
While humanoids are capable of executing impressive tasks like dancing, they currently lack the cognitive ability to function in complex and unpredictable, or “unstructured,” settings. For this to change, they will require algorithms that go beyond a large language model’s grasp of the physical universe, allowing them to manage limbs and grippers to walk and manipulate objects. Some research teams are advancing in the quest to develop more broadly capable AI models for robotic applications.
At the same time, it is becoming apparent that fresh ideas may be essential for the progression of AI. The recent letdown surrounding OpenAI’s GPT-5 reflects a wider acknowledgment that attaining human-like intelligence will call for new avenues of research.
“They’ve asymptoted on GPT-5,” states Tellex. “They need to shift towards the physical world.”
