Humanoid robots are often described as the next great leap in automation — machines that could one day perform surgery, replace factory workers, or even serve as personal assistants in the home. But while artificial intelligence (AI) chatbots like ChatGPT have advanced at remarkable speed, robotics experts caution against assuming humanoid robots will follow the same trajectory.
In two new papers published in Science Robotics, UC Berkeley roboticist Ken Goldberg highlights a critical obstacle: the so-called “100,000-year data gap.” This shortfall in training data means humanoid robots will not acquire real-world skills nearly as quickly as large language models (LLMs) have gained fluency in language.
For the construction and digital engineering sectors, where robotics and automation are seen as key drivers of productivity and safety, understanding this gap is essential for realistic planning and investment.
Why AI chatbots evolved faster than humanoid robots
AI chatbots such as ChatGPT, Claude and Gemini were trained on vast volumes of text from across the internet. This abundance of data enabled large language models to develop fluency in human language, with applications in:
- Customer service automation
- Personal digital assistants
- Content creation and summarisation
- Healthcare triage and support
Goldberg estimates that it would take a human reader around 100,000 years to read through all the text used to train LLMs. Robotics lacks anything close to that scale of training data, particularly for the kind of dexterity and manipulation skills required to perform everyday physical tasks.
In robotics, dexterity is a major challenge. Picking up a wine glass, screwing in a lightbulb, or sorting packages requires spatial awareness, precise motion and feedback control. Unlike processing text, these are embodied tasks that demand physical interaction with the real world — a far more complex challenge for machine learning.
The 100,000-year data gap explained
Goldberg’s research highlights three key reasons why humanoid robots face such a massive data deficit compared to AI chatbots:
1. Limited Real-World Data
Robots require high-quality data from physical environments, but collecting it is slow and expensive. Teleoperation, where humans remotely control robots, generates training data — but one day of operation yields only one day’s worth of data.
2. Challenges in Using Video Data
Some have proposed mining platforms like YouTube for data on human movements. However, translating 2D video into precise 3D motions is technically difficult, and lacks the fine-grained detail needed for robotic manipulation.
3. Simulation Constraints
While simulation environments work for some robotics tasks, such as running or balancing, they fail when it comes to dexterous tasks like plumbing, electrical work, or material handling in construction. Transferring simulated skills into real-world performance remains a barrier.
The debate: data-driven learning vs engineering principles
Goldberg argues that robotics is undergoing a paradigm shift. The field is divided between two schools of thought:
- Data-Driven Robotics: Advocates believe that scaling up data collection and machine learning will eventually deliver humanoid robots capable of human-like performance.
- Engineering-Centric Robotics: Others insist that physics, mathematics and traditional engineering principles remain essential for building functional robots.
The reality, Goldberg suggests, may be a hybrid approach. Robots engineered with baseline functionality can then be deployed in real-world environments to generate the data needed for more advanced machine learning.
Examples include:
- Waymo’s self-driving cars, which continuously collect data from real-world journeys, improving their navigation systems over time.
- Ambi Robotics’ warehouse automation systems, which learn and improve by handling millions of packages in logistics centres.
This “bootstrapping” approach could be the key to bridging the data gap in humanoid robotics.
Implications for digital construction and industry
For the construction sector, humanoid robots are often portrayed as a future solution for labour shortages, site safety, and productivity gains. However, Goldberg’s research indicates that expectations need to be tempered with realism:
- Skilled Trades Remain Secure: Complex physical tasks in construction, such as plumbing, electrical installation, and finishing work, are far beyond the capabilities of current humanoid robots.
- Task-Specific Robots Will Lead First: Specialised robots — for surveying, 3D printing, or automated material handling — are far more likely to deliver value in the near term than general-purpose humanoid machines.
- Data Collection Must Happen on Sites: Just as self-driving cars improve by operating in real traffic, construction robots will need to collect data from real projects to enhance their performance.
The future of humanoid robots
Despite the hype, humanoid robots are unlikely to replace surgeons, factory workers, or construction tradespeople in the next five to ten years. Instead, the near future will likely be shaped by incremental advances in automation, powered by:
- AI-driven design and planning tools
- Digital twins and simulation for site management
- Robotics for repetitive, hazardous, or precision tasks
- On-site data collection to improve long-term performance
Humanoid robots need time, data and engineering
The promise of humanoid robots continues to captivate the public imagination, fuelled by bold predictions from technology leaders. Yet as Ken Goldberg and fellow roboticists argue, the sector faces a fundamental data gap that cannot be closed overnight.
For industries like digital construction, the key is to focus on realistic, incremental automation opportunities while recognising the limitations of humanoid robotics in the near term. Success will depend not only on machine learning but also on the integration of engineering expertise, real-world data collection and practical deployment strategies.
In short: humanoid robots are advancing, but they will not outpace the laws of physics, engineering, and data. For now, construction leaders should continue to invest in specialised robotics, digital construction tools, and AI-driven planning systems that can deliver tangible value today, while keeping a close eye on the longer-term evolution of humanoid machines.
Read next: Digital Twins for carbon and cost reduction in construction – Build in Digital
Are you a building professional? Sign up for a FREE MEMBERSHIP to upload news stories, post job vacancies, and connect with colleagues on our secure social feed.
