Humanoid robots still lack true intelligence, says Meta’s AI chief

Meta’s AI chief Yann LeCun says most companies focus on hardware over cognition, missing breakthroughs needed for truly useful humanoids

The global race to build humanoid robots is gaining momentum — but not necessarily intelligence. Yann LeCun, Chief AI Scientist at Meta and a pioneer in deep learning, has cautioned that most robotics firms today lack the fundamental understanding required to make humanoid robots genuinely useful.

Speaking at the inaugural MIT Generative AI Impact Symposium (MGAIC) in the United States, LeCun said that while dozens of startups are building humanoid robots, few have a roadmap to create systems that can actually think, learn, or adapt. “The big secret of the industry,” he noted, “is that none of those companies has any idea how to make those robots smart enough to be generally useful.”

Hardware progress, cognitive gap

LeCun explained that while robots can be trained for specific, repetitive tasks — such as manufacturing or logistics — achieving broader, domestic utility remains far from reality. “Your household robot will require multiple breakthroughs in AI before it becomes truly functional,” he said, adding that the future of such companies depends on advancing “world model” architectures that enable reasoning and planning like humans.

He also stressed that today’s large language models (LLMs), despite their popularity, are insufficient for powering intelligent robotics. “We’re missing something big,” he said. “AI systems must learn from natural, high-bandwidth sensory data like video. We’ll never reach human-level intelligence by just training on text.”

Towards a world model

According to LeCun, the next leap in AI will depend on developing systems that can interpret and learn from the real world — much like how children absorb knowledge through experience. “A four-year-old has seen as much data through vision as the largest language models trained on all publicly available text,” he pointed out.

LeCun concluded that only world model-based AI systems, which can learn from continuous sensory and visual inputs, will unlock the true potential of humanoid robots — turning them from mechanical marvels into intelligent companions capable of real-world understanding.