The great divide: understanding the debate on the future of humanoid robots
Insights on the latest developments in the world of autonomous mobile robots.
The robotics community stands at a crossroads in 2025, engaged in what UC Berkeley's Ken Goldberg calls a “paradigm shift” — a fundamental disagreement about whether the future of humanoid robots lies in data-driven artificial intelligence or traditional engineering approaches.
This debate, which has captured the attention of leading researchers and tech executives worldwide, centres on a critical question: Will robots achieve human-level capabilities through massive data collection and training, or do they require the structured foundation of physics-based modelling and engineering principles?
Reality or hype?
Tech luminaries have made extraordinary claims about the imminent arrival of general-purpose humanoid robots. Elon Musk predicts that by 2040, there will be 10 billion humanoid robots — more robots than people on Earth. NVIDIA CEO Jensen Huang has similarly bet big on robot-powered factories, while others suggest we'll see robots performing surgery and replacing factory workers within just a few years.[2][3]
However, robotics experts are pushing back against this timeline. As Goldberg emphasizes, while robots are advancing quickly, “it's not going to happen in the next two years, or five years or even 10 years,” Goldberg tells UC Berkeley News. He is co-founder of Ambi Robotics and Jacobi Robotics and William S. Floyd Distinguished Chair of Engineering at UC Berkeley.
The disconnect between public expectations and scientific reality has created what many researchers view as dangerous “humanoid hype” that could lead to a bubble and subsequent backlash.
100,000-year data gap
At the heart of the debate lies what Goldberg terms the “100,000-year data gap”. Large language models that power AI chatbots were trained on internet-scale text data equivalent to what would take a human approximately 100,000 years to read. This massive dataset enabled rapid advances in language fluency and reasoning capabilities.
Robots face a fundamentally different challenge. Unlike text data readily available on the internet, robots require training data that combines video inputs with precise robot motion commands — data that simply doesn't exist today at the required scale. The few approaches to bridge this gap each face significant limitations:
YouTube videos can't provide the detailed 3D motion data needed for dexterous manipulation
Simulation works well for basic locomotion like robot backflips, but fails for complex dexterity tasks like those performed by construction workers or electricians
Teleoperation, where humans control robots remotely, generates data at an impractically slow rate — every eight hours of work yields only eight hours of data
Two schools
One camp, led by researchers like MIT's Daniela Rus and Russ Tedrake, argues that data-driven approaches are essential for robots to function in unpredictable, human-centred environments. Rus emphasizes that “physics gives us clean models for controlled environments, but the moment we step outside, those assumptions collapse”.
This approach focuses on the following:
Multimodal datasets capturing human activities from cooking to object manipulation
Learning from demonstration with robots adapting to variations in real-world scenarios
Scaling effects where increased data leads to emergent robustness and common-sense behaviours
Tedrake demonstrated this with robots learning to slice apples, where each apple's unique characteristics required adaptive responses that emerged naturally from diverse training data.
The opposing camp, including MIT's Leslie Kaelbling, argues for the continuing importance of mathematical models and first principles. Kaelbling states that “data can show us patterns, but models give us understanding. Without models, we risk systems that work, until they suddenly don't”.
Key arguments include:
Safety-critical applications demand deeper understanding than trial-and-error learning
Physics-based models provide essential insights about motion, force, and control that data alone cannot capture
Structured approaches can guide data collection and interpretation more effectively
Despite the debate, significant investment continues flowing into humanoid robotics. Goldman Sachs estimates the market could reach $38 billion by 2035, while other projections suggest $66 billion by 2032. Chinese companies are leading development, with China projected to control nearly one-third of the global market by 2029.
However, fundamental challenges remain. The most basic human capabilities — like dexterity in object manipulation — continue to elude robots. As Goldberg notes, “no robot can pick up a wine glass or change a light bulb”. This reflects Moravec's paradox: tasks that seem simple to humans often prove extraordinarily difficult for machines.
Hybrid solutions
Rather than choosing sides, many experts now advocate for hybrid approaches that combine both paradigms. Goldberg suggests using traditional engineering to create robots functional enough for real-world deployment, which then generates the data needed for improvement — creating a “data flywheel” effect.
This approach is already showing success:
Waymo's self-driving cars collect real-world data daily while operating commercially
Ambi Robotics improves package-sorting robots through continuous data collection in active warehouses
Industrial automation continues advancing through combined model-based control and data-driven adaptation
It takes a village
Contrary to popular fears, Goldberg argues that blue-collar trades are relatively safe from robot displacement. Jobs requiring dexterity and adaptability — construction workers, plumbers, electricians — remain far beyond current robot capabilities. Instead, routine cognitive tasks like form-filling and certain aspects of customer service face more immediate automation risk.
The debate over humanoid robots reflects a healthy scientific process where different approaches compete and ultimately combine to drive progress. While the timeline for general-purpose humanoid robots remains longer than tech leaders suggest, the field is advancing through both data-driven learning and engineering innovation.
The key insight from this debate is that neither approach alone is sufficient. The future likely belongs to hybrid systems that leverage the pattern recognition capabilities of large-scale data while respecting the constraints and insights provided by physics-based models.
As the robotics community continues this paradigm shift, managing public expectations while pursuing rigorous research will be crucial to avoiding the boom-and-bust cycles that have historically plagued the field.
The humanoid robot revolution may not arrive in the next few years, but the foundations being laid today — through both camps in this great debate — are building toward a future where robots can safely and effectively operate alongside humans in the complex, unpredictable real world.
