As factories finalize 2026 budgets, a costly choice looms: invest in human-shaped machines (humanoid robots) capturing headlines, or purpose-built systems capturing ROI.
When astronauts perform spacewalks outside the International Space Station, they face a persistent problem: not enough hands. NASA’s solution wasn’t humanoid robots. It was Canadarm2 and Dextre—a two-armed manipulator with specialized end effectors that works 24 hours without rest. These systems extend human capability rather than replicate human form. Yet on Earth, billions in venture capital flood into companies building bipedal humanoid robots that look like us, move like us, and fail like us.
A general view shows humanoid robots of XIngdong L7 working on an assembly line at ROBOTERA at the Shanghai New Expo Center during the opening day of the World Artificial Intelligence Conference (WAIC) 2025 in Shanghai, China, on July 26, 2025. (Photo by Ying Tang/NurPhoto via Getty Images)
NurPhoto via Getty Images
At CES 2025, NVIDIA CEO Jensen Huang declared physical AI “the next big thing,” positioning humanoid robots as the future of the $50 trillion manufacturing and logistics industries. Goldman Sachs projects a $154 billion market by 2035. Fellow Forbes Contributors seem enthusiastic; Ethan Karp sees opportunities, Cornelia Walther calls them inevitable, fellow futurist Bernard Marr predicts a robotic culture shock, although Dev Patnaik worries about social permission, at least in the home. The Manufacturing Leadership Council reports 22% of manufacturers plan to deploy humanoid robots within two years. But the International Federation of Robotics recently questioned whether humanoid robots represent “an economically viable and scalable business case for industrial applications.”
The Reality Check From Robotics Pioneers
Rodney Brooks—co-founder of iRobot and MIT professor emeritus—delivered the bluntest assessment in September 2025: humanoid robots pursuing dexterity through video-based learning represent “pure fantasy thinking.” Brooks identifies the hardware deficit: human hands contain approximately 17,000 specialized tactile receptors; humanoid robots have effectively zero. The infrastructure for collecting high-fidelity haptic data—the ImageNet equivalent for touch—doesn’t exist. Every humanoid robot failure cascades into downtime, liability exposure, and retraining costs vendors exclude from glossy demos.
The challenge goes beyond sensors. As Meta’s Chief AI Scientist Yann LeCun argues, true machine intelligence requires building internal “world models” that predict physical consequences. A welding robot needs to model a millimeter-scale workspace. Humanoid robots attempting general-purpose manipulation must model the entire world. They’re trying to master an entire library of physics when factories need a single, perfectly executed chapter.
Then there’s safety. Brooks instituted a personal “three-meter rule” after witnessing a full-size bipedal humanoid robot fall at close range. When they fail—and they will—the consequences scale exponentially. A robot twice current size packs eight times the destructive kinetic energy. Regulators have no established safety envelopes for bipedal humanoid robots moving unpredictably through shared workspaces, creating compliance friction that delays deployment.
The Form Factor Fallacy of Humanoid Robots
The assumption driving humanoid robot investment is seductive: humans designed factories for human bodies, so human-shaped robots slot directly into existing infrastructure. But this inverts the design challenge. We need robots that accomplish tasks humans struggle with, in conditions humans find difficult, with performance humans can’t match.
Consider deployed alternatives: Boston Dynamics’ Spot quadruped operates in oil platforms and automotive lines with stability bipedal humanoid robots can’t achieve. Universal Robots’ UR15 demonstrates AI-driven motion without humanoid bodies. Dextre, the Canadian robot that maintains the International Space Station since 2008 uses two seven-joint arms—one for stability, one for work—without wasting energy on bipedal balance.
Toyota’s robotics philosophy prioritizes incremental, purpose-built systems over humanoid robot moonshots. While Hyundai has announced future humanoid robot experiments, their Factory of the Future production systems today rely on autonomous mobile robots and collaborative arms—proven technologies delivering ROI now, not promises for later. These aren’t companies afraid of automation; they understand the difference between PR and production economics.
The economics are stark. Major humanoid robot vendors avoid discussing costs, but industry estimates place unit costs at $120,000-$200,000. Collaborative arms and autonomous mobile robots cost 40-60% less with established maintenance supply chains. Specialized systems achieve 95%+ uptime while humanoid robots face mechanical stress from bipedal locomotion that degrades components cyclically. Typical AMRs run 20-22 hours daily with predictable maintenance. No humanoid robots on the market achieve anything close.
Manufacturing Needs Non-Humanoid Robots
Physical AI represents genuine progress—just not through humanoid robots. The World Economic Forum analysis shows Amazon scales robotic arms and mobile robots across 300+ fulfillment centers, achieving 25% efficiency gains. Foxconn uses AI and digital twins for precision tasks, cutting deployment time 40%. The breakthrough isn’t humanoid robot form—it’s intelligence applied to task-specific systems.
Before committing capital to any physical AI system, executives should apply three tests: Does it extend human capability or reproduce it? Is the form factor minimum necessary? Can you quantify ROI within 18 months? If humanoid robot form is required to answer yes to all three, buy it. Otherwise, you’re paying what I call the Generalization Tax—the massive overhead of building humanoid robots that do 1,000 tasks poorly instead of 10 tasks perfectly. If a system fails any test, it’s speculation—not a manufacturing asset.
The Species We Already Have
The humanoid robot obsession isn’t irrational. After large language models plateaued in mid-2025, the AI industry correctly identified that physical intelligence requires embodied learning. The error is assuming humanoid robot embodiment is the only path.
But there’s a deeper fallacy: Earth hosts 8 billion humans who reproduce reliably, train adaptably, and excel at general-purpose problem-solving. If the goal is beings that look human and think human, we have abundant supply. What manufacturing needs are systems that perform tasks humans cannot do, cannot do safely, or cannot do economically at scale. We don’t need artificial humans. We need superhuman specialists.
This is the key psychological dimension: purpose-built machines are non-threatening amplifiers. Spot and Canadarm2 remain “suitably inferior” in their obvious machine-ness, amplifying human capability without threatening our sense of specialness. Humanoid robots that approach but never achieve human capability force us to question what makes us special if it can be mechanically reproduced—a feeling that degrades workplace trust and collaboration.
The Dynasphere, 1938. The Dynasphere was a monowheel vehicle patented by JA Purves in 1930. Churchman’s cigarette card, from a series titled Modern Wonders [WA & AC Churchman, Great Britain & Ireland, 1938]. (Photo by The Print Collector/Getty Images)
Getty Images
Consider what this reveals: the most transformative tools in history—the wheel, the lever, the computer—succeeded precisely because they didn’t replicate human form. They extended human capability into domains our biology couldn’t reach.
Manufacturing executives must recognize this isn’t just ROI calculation. It’s a choice about what role humans play in production. Do we design systems that let humans manage complexity with superhuman tools? Or chase human-shaped machines that promise to make humans optional?
Every dollar spent chasing humanoid robot generalization is a dollar not spent eliminating bottlenecks, removing hazards, or accelerating cycle time. As 2026 budgets lock in, manufacturers face choices with decade-long consequences. Choose tools that extend reach, not toys that reproduce form. In manufacturing, imitation isn’t innovation—it’s an avoidable cost.
We don’t need more humans. We have billions. We need machines designed for what humans can’t do—not humanoid robots dressed up to look like what we can. The future belongs to machines that make humans more capable, not more replaceable. Your investment decision reveals which future you’re building.
Trond Undheim is author of “The Platinum Workforce” (Anthem Press, 2025) and former Research Scholar at Stanford’s Center for International Security and Cooperation, and a humanoid robots skeptic.