The short answer: Robots do less than 1% of the world's work because the technology is still too hard to use, too hard to certify as safe, and too disconnected from how humans actually describe tasks. The bottleneck is not intelligence. It is hands, force control, and the semantic gap between what a worker means and what a servo motor executes. That is the assessment of Klas Nilsson, who built one of the first digital robot motion control systems at ABB in the 1980s, became a robotics professor at Lund University in Sweden, and founded Cognibotics.
On this week's episode of Unscripted, Nilsson separated 40 years of engineering reality from the current humanoid robot hype. Here is what he sees that most of the industry does not.
Everyone obsesses over robots that walk. Nilsson deliberately chose the other end: arms, grippers, and manipulation. Mobility is crowded and progressing well. Hands are not.
The human hand packs enormous strength into very little weight, with dozens of degrees of freedom. It also relaxes when it moves fast, so bumping into something does no damage. No robotic hand does that today. Chinese manufacturers are miniaturizing motors directly into fingers. Tesla is betting on tendon-driven designs. Nilsson's verdict: none of the current approaches will win. Variable stiffness in hardware remains unsolved, and the materials are still too heavy.
His nod goes to Boston Dynamics for pragmatism: ship a simpler hand now, build better hands later.
Nilsson admitted he recently changed his mind on this. He could never see how AI action tokens could coexist with a factory worker who needs to adjust a robot's motion slightly, without launching a retraining project.
The answer he landed on: layers. Keep LLMs at the top, where humans define and adjust tasks in natural language. Below that, digital twins and industrial software handle programming, monitoring, and supervision. At the bottom, deterministic servo control does the physics. Cognibotics builds the connective tissue between those layers, including a Julia-based real-time language called Juliet.
The insight that matters for any leader evaluating AI: what a human means by "force" and what a servo controller means by force are different things. Collapsing those layers into one giant model is how demos get made. Separating them is how factories run.
Nearly every industrial robot today is programmed by position and orientation only. Nobody tells the robot how hard to push, what torque is acceptable, or how much is too much. That knowledge stays in the heads of the humans who teach them.
"Of course they are stupid," Nilsson said. Until robots can be given expressive instructions about force, uncertainty, and monitoring, they cannot take on the craftsman-like work that represents most human labor. Solve that in manufacturing first, he argues, and humanoids follow.
Nilsson's forecast is specific. The hype phase ends within roughly 2 years, replaced by a realism phase where results from actual factories and logistics centers matter more than YouTube videos. Real breakthroughs in general-purpose humanoid robots take 7 to 10 years. Home robots for ordinary households take longer still, with wealthy early adopters getting useful units in 7 or 8 years.
He also sees billions in current investment flowing into approaches that cannot work commercially or technically. His example: models that start from geometry rather than physics, force, and elasticity look scalable but break down on craftsman-like motion. The coming correction will redirect capital toward what actually functions on a factory floor.
While AI threatens graphic designers and legal staff, Nilsson sees the opposite for trade labor. Future robots will learn by haptic teaching: a carpenter or bricklayer physically showing the motion, holding the tool, guiding by force. Skilled workers become robot teachers, and can sell the services their trained robots deliver.
His broader frame is worth repeating: LLMs are predictors, not intelligence. A robot programmed with "empathy" is still a machine with none, functionally a psychopath. Robots should remain machines that serve humans. In his words, humans should do human things.
Why do robots perform less than 1% of the world's work? Because the technology is not scalable or easy enough to use. Hardware is heavy and hard to certify as safe at full speed near humans. Software makes it difficult to define tasks, adjust behavior, and express force requirements. The barriers are technical, not economic demand.
Will LLMs solve robotics? No. LLMs work at the task-definition level: telling a robot what to do in human terms. Motion control, force, and safety live in lower system layers that need deterministic, real-time engineering. The unsolved challenge is linking the two without semantic gaps.
When will humanoid robots be in homes? Nilsson estimates 7 to 8 years before wealthy early adopters have useful home humanoids, and 10+ years before general availability. Manufacturing and logistics adoption comes first.
Will robots take skilled trade jobs? Nilsson argues the reverse: skilled touch labor becomes more valuable because craftsmen will teach robots by demonstration and force guidance, creating new robot-teacher roles and service markets.
What is Cognibotics? A Swedish deep-tech robotics company spun out of Lund University, focused on calibration, accuracy, motion control, and the software foundations that make robots precise and easier to program.
Unscripted with Jeff Pedowitz features honest conversations about AI with leaders across security, science, business, medicine, technology, and the arts. Listen to the full episode with Klas Nilsson wherever you get your podcasts.
The Pedowitz Group | pedowitzgroup.com