A renewed concentrate on robots would counsel that OpenAI believes reaching synthetic normal intelligence (AGI)—AI that exceeds human intelligence—might require growing algorithms which might be able to interacting with the bodily world.
OpenAI did notable robotics analysis in its early years, together with growing an algorithm able to fixing a Rubik’s dice utilizing a humanlike hand in 2019. The corporate shuttered its robotics effort in 2021, nonetheless, to concentrate on algorithms together with the big language fashions which have pushed current breakthroughs akin to ChatGPT. OpenAI restarted work on robots last year, and The Information reported in December 2024 that the corporate was contemplating growing its personal humanoid robots.
Stefanie Tellex, a roboticist at Brown College, says that constructing simpler robots will contain designing and coaching AI fashions able to “processing high-frame-rate, high-dimensional perceptual enter, and producing high-frame-rate, high-dimensional bodily outputs”—that means fashions that may see and act with excessive constancy. Tellex isn’t accustomed to OpenAI’s plans particularly, nonetheless.
Regardless of already having industry-leading fashions for dialog, reasoning, coding, and picture and video technology, OpenAI shall be racing a sequence of sturdy opponents because it seeks to develop the algorithms for extra succesful humanoid robots. A handful of humanoid startups, together with Figure, Agility, and Apptronik, have emerged over the previous few years, and a few main AI corporations, together with Tesla and Google, are additionally investing in growing and testing humanoids. “I don’t see them having any magical benefit over anybody else,” says Tellex.
Humanoids have gotten more and more standard because the {hardware} and software program wanted to construct functioning prototypes turns into extra widespread. Whereas humanoids are nonetheless costly and tough to develop, new sorts of motors and different parts have made it cheaper and simpler to place collectively functioning methods. Software program akin to Nvidia’s Isaac robotic growth platform have additionally made it less complicated to jot down the code wanted to regulate and prepare humanoid methods.
Humanoid hype can also be constructing. Enterprise capitalists have invested more than $5 billion in humanoid startups because the begin of 2024. Morgan Stanley reckons that the humanoid {industry} could be worth $5 trillion by 2050.
Whereas humanoids can carry out spectacular feats like dancing, they nonetheless lack the intelligence required to function in complicated and unpredictable, or “unstructured,” environments. To accumulate this, they may want algorithms that transcend a big language mannequin’s understanding of the bodily world. These methods should be capable of management limbs and grippers so as to stroll and manipulate bodily gadgets. Some analysis teams are beginning to reveal progress in growing extra typically succesful AI fashions for robots.
On the similar time, it’s changing into more and more clear that new concepts could also be wanted to push AI ahead. The current disappointment of OpenAI’s GPT-5 is a part of a broader realization that reaching humanlike intelligence would require new avenues of analysis.
“They’ve asymptoted on GPT-5,” says Tellex. “They should transfer in the direction of the bodily world.”
