If your email corresponds to an account, you will receive a reset link. The humanoid robotics sector, supported by billions in investments, is said to be facing a bubble risk according to prominent figures like Rodney Brooks (co-founder of iRobot) and Yann LeCun (Meta’s Chief AI Scientist). They highlight the current approaches’ inability, especially in terms of dexterity and general intelligence, to justify the promises made by companies such as Tesla and Figure.
The race to develop autonomous and versatile humanoid robots is thriving. Companies like Figure, recently valued at an “astonishing” $39 billion after a billion-dollar fundraising, and Tesla with its Optimus project, are nurturing ambitious goals. Figure’s CEO, Mike Cagney, and Elon Musk, promise a significant economic impact within five years. However, two highly regarded minds in the field, roboticist Rodney Brooks and Meta’s Chief AI Scientist, Yann LeCun, have cast doubt on this financial optimism.
Brooks, a renowned roboticist with decades at MIT, co-authored an article explaining “Why today’s humanoids won’t learn dexterity.” His assessment is clear: the hundreds of millions, even billions, invested by venture capitalists and tech giants for their training is spent on an approach that cannot succeed. Brooks believes that achieving human-like dexterity in the coming decades is “pure fantasy.”
The core issue lies in the hands. Human hands have around 17,000 specialized touch receptors. According to Brooks, no current robotic technology comes close to this capability. While machine learning has transformed speech recognition and image processing through decades of specific data, there’s no equivalent “tradition” for the touch data robots would need. Attempts by some companies, like Figure or Tesla, to teach robots dexterity by showing them videos of humans performing tasks are particularly criticized by iRobot’s co-founder. He points out that efforts to build human-like hands, even though ongoing for decades, have not yet solved this fundamental bottleneck related to acquiring rich sensory data.
On the other hand, Turing Award winner and deep learning pioneer Yann LeCun questions the machines’ intelligence. Meta’s Chief Scientist warned at MIT’s inaugural Impact of Generative AI Symposium that the “big secret of the industry” is that none of these companies has a clue on how to make these robots smart enough to be “generally useful.” While robots can be trained for specific tasks, like in manufacturing, a domestic robot will require major AI breakthroughs. LeCun believes that current large language models are not the solution and emphasizes that training solely on text will never achieve human-level intelligence.
To move forward, the focus is on what are called world models. These AI systems learn to understand the physical world from sensory data (video) by predicting the future state of the world after an imagined action by the agent. LeCun, researching architectures like V-JEPA, is convinced that these models are key for robots to perform “zero-shot” tasks without training.
Beyond intelligence and dexterity, Rodney Brooks raises a often-overlooked point: safety. Bipedal humanoid robots, due to the massive energy required to stand and walk, pose a significant danger in case of falls. This physical issue, combined with software challenges, leads Brooks to predict that successful humanoids in fifteen years will bear little resemblance to current anthropomorphic models. They will likely be equipped with wheels, multiple arms, and specialized sensors, abandoning the human form for efficiency reasons.
The alert raised by Brooks and LeCun forces the industry to address a fundamental question: can massive funding of expensive training experiments truly lead to scalable mass production without first addressing fundamental AI bottlenecks? The debate rages on, and the next five years, as set by some entrepreneurs, will be the judge of the current humanoid form’s viability.
About the Author
