The venture capital taps are wide open, YouTube demos are racking up views in the millions, and the quarterly promises are growing ever more audacious. After watching large language models like OpenAI’s ChatGPT seemingly conquer the digital realm overnight, the tech world is asking the same multibillion-pound question: when will robotics finally have its “ChatGPT moment”?
According to two men who have spent years in the trenches of embodied AI, the answer is blunt: it won’t. And we would be wise to listen. Jonathan W. Hurst, co-founder of Agility Robotics (the team behind the Digit humanoid), and Hans Peter Brøndmo, who headed the Everyday Robots moonshot at Google X, have penned a necessary reality check. They are here to pour a bucket of industrial-grade coolant over the hype train, arguing that no single, miraculous AI breakthrough is going to save the day. Instead, the path to a world of functional robots is paved with gruelling, incremental, and deeply unsexy engineering.
With venture funding for robotics hitting $6.1 billion in 2024—up from $5.1 billion in 2023—the stakes are stratospheric. But as Hurst and Brøndmo argue, the chasm between a flashy demo and a commercially viable, safe, and reliable machine remains vast.
The Grand Illusion: Deconstructing the YouTube Demo
We’ve all seen them: the viral clips of humanoid robots executing flawless backflips, performing synchronised dances, or engaging in choreographed martial arts. The latest sensation featured Unitree Robotics’ humanoids performing kung fu at the 2026 Spring Festival Gala in China—an impressive display of coordination just inches away from child performers.
Hurst and Brøndmo are quick to point out what industry insiders have known for years: “never trust a YouTube robot video.” These performances, while technically brilliant, are the robotic equivalent of a meticulously rehearsed West End show. They showcase remarkable low-level motor control, but their level of autonomy is closer to a pre-programmed assembly-line arm than a sentient machine. The real world—messy, unpredictable, and stubbornly unscripted—is a different beast entirely. It is a classic manifestation of Moravec’s paradox: tasks that are trivial for a human, such as navigating a cluttered hallway, are monumentally difficult for a robot, while complex calculations are child’s play.
Data is a Herculean, Unsolved Challenge
LLMs had a massive head start: they were trained on the internet, a colossal, human-generated archive of text and images. Robots enjoy no such luxury. To learn, a robot requires data from the physical world—a high-dimensional space where variables range from joint angles and force feedback to shifting light and the erratic movements of people.
The scale of this challenge is staggering. At Everyday Robots, Brøndmo’s team ran 240 million simulated robot instances in 2022 just to train a model to sort rubbish with something approaching competence. And that was for a single, relatively basic skill set. Now, multiply that effort across the near-infinite number of tasks we expect general-purpose robots to master. It is a data collection problem of a completely different magnitude, and one that remains largely unsolved. Indeed, the project itself was shuttered in early 2023 as part of wider cost-cutting measures at parent company Alphabet.
There Will Be No Single “Robot AI”
The notion of a single, monolithic AI model that can pilot any robot—whether it’s wheeled, legged, flying, or swimming—is pure science fiction. The physical realities of different forms and environments are simply too varied. The authors contend that the winning architecture will be what they term “agentic AI.” This involves high-level coordinating models that can reason, plan, and delegate tasks to a suite of smaller, specialised AI tools. One model might handle bipedal locomotion, another fine-grained manipulation, while a third is dedicated solely to safe human-robot interaction.
This modular approach, they argue, will lead to a “Cambrian explosion” of useful, intelligent machines. It won’t be a single Big Bang, but a blossoming of diverse, specialised capabilities that, when orchestrated correctly, create a truly competent machine.
Hardware is Still Agonisingly Hard
For all the focus on AI, a robot remains a physical object. And the hardware—specifically the components that allow it to interact with the world—remains a major bottleneck. Most industrial robots rely on rigid, powerful actuators that are fantastic for precision in a safety cage but disastrous in a human environment. A simple, accidental collision could be catastrophic.
Humans, by contrast, are “compliant.” We use touch and force feedback constantly, whether we’re jiggling a key into a lock or leaning on a kitchen counter. For robots to achieve this, they need a new class of actuators that are sensitive, compliant, and force-aware. While these exist in research labs, they are not yet available at the scale, cost, or reliability required for mass deployment. The most brilliant AI in the world is effectively redundant if its body is a clumsy, dangerous liability.
Real Value Comes from Solving “Easy” Problems
The final, and perhaps most vital, truth is that real-world value doesn’t come from backflips. It comes from reliably performing the mundane, repetitive, and often physically taxing jobs that humans would rather avoid. This is the business end of the industry—where the robot’s feet meet the warehouse floor.
Both authors speak from hard-won experience. When Agility Robotics began deploying Digit at customer sites with partners like GXO Logistics, they quickly realised their first major hurdle wasn’t task performance—it was safety. This led to a multi-year engineering effort to redesign the robot for safe operation alongside people. Similarly, the Everyday Robots team at Google learned first-hand how chaotic a seemingly simple environment like an office canteen is for a robot trying to clear tables.
This real-world exposure is the only way forward. It informs the AI architecture, highlights hardware flaws, and grounds ambitious roadmaps in the harsh reality of customer needs. There is no silver-bullet algorithm or dataset that can substitute for the slow, painful, and expensive process of deploying robots, watching them fail, and meticulously engineering solutions. The future of robotics is coming, but it will arrive one deliberate, well-engineered step at a time.