Humanoid robots are advancing but remain far behind the rapid progress of large language models, according to new research from the University of California, Berkeley. In two papers published in Science Robotics, UC Berkeley professor Ken Goldberg argues that robotics faces what he calls a “100,000-year data gap,” which will slow the development of real-world robotic skills compared with the fluency achieved by AI chatbots.
Goldberg explains that large language models were trained on vast amounts of text data, the equivalent of what would take a human about 100,000 years to read. By contrast, no comparable dataset exists for robots. While some researchers propose training from online video or simulations, Goldberg notes that current methods fail to capture the fine-grained physical dexterity required for everyday tasks, such as picking up a glass or changing a light bulb. Efforts to bridge the gap through teleoperation, in which humans remotely control robots, generate data at a much slower rate.
The two papers highlight a growing divide within robotics. One camp supports data-driven methods modeled after the training of language models, while another emphasizes traditional engineering approaches based on physics and mathematical modeling. Goldberg suggests that integrating both approaches may be necessary, enabling robots to function well enough in real-world environments to generate their own data over time.
Examples of this incremental approach include self-driving cars from Waymo, which improve by collecting data during real-world operation, and Ambi Robotics, which deploys warehouse robots that learn through continued use.
Goldberg remains skeptical of predictions by technology executives, such as Elon Musk and Jensen Huang, who claim humanoid robots will soon surpass human surgeons or function as household assistants. He argues that dexterity challenges and the data shortage mean such capabilities are unlikely to emerge in the next decade.
On the question of jobs, Goldberg sees blue-collar roles in the trades as relatively insulated from automation due to their reliance on physical dexterity, while some white-collar tasks, such as form processing or medical image analysis, are more vulnerable. He also notes that while AI may support functions like radiology or customer service, human interaction remains essential in situations that require empathy or sensitive communication.
Goldberg concludes that while robotics will continue to advance, expectations should be tempered to avoid a speculative bubble that could lead to disillusionment in the field.
