Rodney Brooks Has Been Right About Humanoid Hype for 40 Years. Is This the Time He Is Wrong?

In January 2026, Rodney Brooks updated his annual predictions scorecard. The post, published on his personal blog at rodneybrooks.com, runs several thousand words and methodically evaluates dozens of predictions he first made in 2018. On humanoid robots, Brooks is characteristically blunt. CEO claims about imminent mass deployment of humanoids, he writes, “are just not plausible.” He applies what he calls “the date tethering test” to every bold claim about robots transforming labor markets and finds most of them wanting.

This is not new territory for Brooks. He has been calling out robotics hype since before most of today’s humanoid robot companies existed. Since before most of their founders were born. He has been doing it with a consistency and specificity that makes him uniquely valuable in a field drowning in breathless press releases and investor decks. And he has been right, over and over again, for four decades.

But here is the question that makes this moment different from every other hype cycle Brooks has navigated. As of March 2026, more than 15,000 humanoid robots have actually shipped. Not prototypes. Not demos. Shipped units, deployed in factories, warehouses, research labs, and increasingly in commercial settings. Goldman Sachs has revised its humanoid robot market forecast from $6 billion to $38 billion. China alone has seven companies shipping humanoids in volume. The global race is not theoretical. It is underway.

So the question becomes: Is the man who has been right about robotics hype for 40 years finally facing a revolution he is underestimating?

To answer that question properly, we need to understand who Rodney Brooks is, how he thinks, what he has gotten right, and where his framework might have blind spots. This is not a takedown. Brooks deserves far more respect than most people in the robotics industry give him. But it is also not hagiography. Even the best frameworks have limits.

The man who changed how we think about robots

Rodney Allen Brooks arrived at MIT in 1984 as an Australian computer scientist with radical ideas about artificial intelligence. At the time, mainstream AI research was dominated by what is now called Good Old-Fashioned AI, or GOFAI: systems that built internal models of the world and reasoned about them symbolically. Robots in that paradigm needed detailed maps, explicit planning algorithms, and carefully structured knowledge bases.

Brooks thought this was fundamentally wrong.

His 1986 paper “A Robust Layered Control System for a Mobile Robot” introduced what he called the subsumption architecture. The core insight was that intelligent behavior could emerge from layers of simple reactive systems, each responding directly to sensor input, without any central model of the world. His follow-up papers, including the provocatively titled “Intelligence Without Representation” (1991) and “Elephants Don’t Play Chess” (1990), laid out a vision of robotics that was bottom-up rather than top-down. Instead of trying to model the entire world and then plan within that model, Brooks argued, you should build creatures that react to the world as they find it.

This was not just an academic exercise. The subsumption architecture directly led to the Roomba.

In 1990, Brooks co-founded iRobot with his students Colin Angle and Helen Greiner. The company spent a decade doing military and research robotics before launching the Roomba in 2002. The Roomba was, in many ways, a vindication of Brooks’ philosophical approach. It did not build maps of your home (at least the early models did not). It bounced off walls, sensed drops, followed simple behavioral rules, and cleaned your floor effectively enough that iRobot eventually sold over 40 million units.

Brooks directed MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) from 2003 to 2007, one of the most prestigious positions in the field. Then he left MIT and iRobot to co-found Rethink Robotics in 2008, a company aimed at building collaborative industrial robots. Rethink produced Baxter and Sawyer, robots designed to work alongside humans in manufacturing environments.

Rethink Robotics failed. It shut down in 2018 after raising over $150 million. The robots were too slow, too imprecise, and too limited for the tasks manufacturers actually needed done. This failure is important context for understanding Brooks as a thinker. He is not someone who just criticizes from the sidelines. He built a company, raised venture capital, tried to commercialize robots for real-world work, and learned firsthand how brutally hard it is to make robots that do useful things in unstructured environments.

That failure, far from disqualifying him, makes his skepticism more credible. He knows what it takes because he tried and fell short.

The prediction methodology

What distinguishes Brooks from the average tech skeptic is rigor. Most people who push back on robotics hype do so with vague hand-waving about how “it is harder than it looks.” Brooks does something much more specific.

In 2018, he published “My Dated Predictions” on his blog, a detailed list of technology predictions with specific dates and confidence levels. He then committed to updating the scorecard every January 1st, publicly grading himself against his own past predictions. He has kept this commitment without fail for eight years now.

His methodology involves several key principles.

Date tethering. Every prediction must include a specific date by which it will or will not come true. Brooks argues that predictions without dates are meaningless because they can never be falsified. “By 2025, we will have self-driving cars” is a prediction. “Self-driving cars are coming” is not.

The 3x rule for Musk timelines. Brooks has observed that Elon Musk’s technology predictions consistently follow a pattern: take the predicted delivery date, calculate the time from announcement to predicted date, multiply by three, and add that to the announcement date. The result is far more accurate than Musk’s original prediction. This heuristic has proven remarkably consistent across Full Self-Driving, the Tesla Semi, and Optimus.

The installation base test. When someone claims a technology will be ubiquitous by year X, Brooks asks: what is the current installed base? What is the growth rate? Is it mathematically possible to go from the current base to the predicted base in the claimed timeframe, given manufacturing constraints? In almost every case, the answer is no.

The deployment reality check. Brooks distinguishes sharply between demonstrations and deployments. A robot doing something impressive in a lab or on a stage is not evidence that the technology is ready for the real world. The gap between “works in a demo” and “works reliably in production, 24/7, for months” is, in Brooks’ experience, consistently underestimated by roughly an order of magnitude.

The track record: What Brooks got right

This is the part that should make anyone who dismisses Brooks uncomfortable. His track record is extraordinary. Let us walk through the major hype cycles he has called correctly.

The self-driving car prediction is perhaps his most impressive call. When Musk, Waymo, and others were predicting fully autonomous vehicles by 2017-2020, Brooks publicly stated the timelines were far too aggressive. He was specific: he predicted that a self-driving car would not be able to drive from the west coast to the east coast of the United States, under all conditions, without human intervention, by January 2030. Almost a decade later, that prediction is still looking solid.

His Tesla-specific predictions have been equally sharp. When Musk announced Tesla Bot at AI Day 2021, Brooks immediately applied his 3x rule. Musk said a prototype would be ready “next year” (2022) and production would begin within 3-5 years (2025-2027). Brooks’ 3x calculation suggested 2024-2025 for a real prototype and 2030-2033 for meaningful production. The prototype part was roughly correct. The production timeline is still playing out.

The 2026 scorecard: What Brooks says now

Brooks’ January 2026 predictions scorecard is detailed and worth reading in full. But the key claims about humanoid robots can be summarized as follows.

First, he argues that CEO timelines for mass deployment of humanoid robots are not plausible. When companies claim they will ship thousands or tens of thousands of humanoids within one to two years, Brooks sees the same pattern he has seen in every previous hype cycle: ambitious projections disconnected from manufacturing reality, deployment complexity, and the sheer difficulty of making robots work reliably in diverse real-world settings.

Second, he remains skeptical about the economics. Building humanoid robots at scale requires solving actuator costs, battery costs, sensor integration, and software reliability simultaneously. No one, Brooks argues, has demonstrated that they can do all of these things at a price point that makes economic sense compared to existing automation or human labor.

Third, he is skeptical about generalization. The demo videos look impressive. A robot folding laundry, sorting parts, carrying trays. But Brooks asks: can that same robot do all of those tasks without being reprogrammed? Can it handle the edge cases? What happens when the laundry is wet, the parts are a different size, the tray is heavier than expected? The gap between a scripted demo and robust general-purpose operation is, in his view, still enormous.

Fourth, and this is perhaps his most nuanced point, Brooks does not deny that humanoid robots will eventually be a significant market. He denies that the timeline being promoted by companies and investors is realistic. The difference between “humanoid robots will transform manufacturing by 2030” and “humanoid robots will transform manufacturing by 2040” is the difference between a revolution and an evolution. Brooks believes we are in the evolution.

Where the data challenges Brooks

Now we get to the hard part. Because while Brooks’ methodology has served him extremely well for 40 years, the current moment in humanoid robotics has some features that do not map cleanly onto previous hype cycles.

The shipment numbers are real. This is the single most important difference between 2026 and every previous humanoid hype cycle. In the Honda ASIMO era, total humanoid shipments worldwide were essentially zero. In the 2021-2022 humanoid hype wave, shipments were in the low hundreds. As of March 2026, the total is over 15,200 units. That is not a demo. That is not a press release. That is a supply chain producing and delivering robots at increasing scale.

China changes the manufacturing equation. Previous robotics hype cycles took place primarily in the context of American and Japanese corporate R&D, where manufacturing scale-up was slow and expensive. The current cycle is being driven substantially by Chinese companies with access to the same supply chain ecosystem that produces consumer electronics at scale. Unitree, AgiBot, UBTECH, Leju, Engine AI, and Fourier Intelligence are all Chinese companies. Together they account for roughly 85% of global humanoid shipments. They operate in an environment where component costs are lower, manufacturing iteration is faster, and government industrial policy actively subsidizes humanoid robot development.

Brooks’ installation base test still applies, and 15,200 units is a very long way from the “millions” that some forecasters predict by 2030. But the growth trajectory is steeper than anything we have seen in previous humanoid robot cycles. And the manufacturing infrastructure to sustain that growth already exists in Shenzhen, Hangzhou, and Shanghai.

The price curve is moving faster than expected. Unitree’s G1 humanoid sells for roughly $16,000. That is less than a new car. It is less than a year’s salary for a warehouse worker in most developed countries. The economics of humanoid robots do not need to make sense at $150,000 per unit. They need to make sense at $15,000-$30,000 per unit, and at least one company is already in that range. Brooks’ economic skepticism was calibrated against robots costing hundreds of thousands of dollars. The price curve has moved underneath him.

Goldman Sachs revised their forecast 6x. In early 2024, Goldman Sachs projected a $6 billion humanoid robot market by 2035. By late 2024, they had revised that to $38 billion. A 6x revision in less than twelve months from one of the world’s most conservative institutional forecasters suggests that something fundamental has shifted in the underlying data, not just in the hype.

The use cases are narrower and more practical than previous cycles promised. This is a subtle but important point. Previous humanoid hype cycles promised general-purpose household robots that would cook, clean, and care for the elderly. The current cycle is being driven by much more specific use cases: factory logistics, parts sorting, palletizing, and quality inspection. These tasks are repetitive, structured enough to be tractable, and high-volume enough to justify the investment. Brooks’ critique of overpromising on generality is valid. But the industry has gotten smarter about targeting achievable use cases first.

Where the data supports Brooks

It would be a mistake, however, to conclude that Brooks is simply wrong. Several aspects of the current landscape validate his core concerns.

The “millions by 2030” projections are still not plausible. Some forecasters and company executives have predicted millions of humanoid robots deployed by 2030. Brooks’ installation base test demolishes this. Going from 15,200 units in March 2026 to one million units by December 2030 would require a compound annual growth rate of roughly 200%. That would be unprecedented for a physical product at this price point. For comparison, industrial robots (traditional articulated arms, not humanoids) took over 20 years to go from 15,000 cumulative units to one million. Even with China’s manufacturing advantages, the 2030 million-unit target requires a growth rate that has no historical precedent.

Demo-to-deployment gaps persist. The most impressive humanoid robot demos remain substantially ahead of what is being deployed in production. Videos of robots folding laundry, making coffee, and performing complex manipulation tasks are almost always shot under controlled conditions with significant human oversight. The robots actually deployed in factories are doing much simpler tasks: carrying trays, sorting parts, moving boxes. Brooks’ point about the gap between demos and reliable production operation remains valid.

Reliability data is scarce. How many hours of continuous operation can a current humanoid robot sustain before requiring maintenance or intervention? The honest answer is that most companies are not publishing this data. Industrial robots from companies like FANUC and ABB routinely operate for 80,000+ hours with minimal maintenance. Humanoid robots, with their complex bipedal locomotion systems, dexterous hands, and high-DoF actuators, almost certainly cannot match that yet. Until reliability data is transparent, the true cost of ownership remains unknown.

Software is the bottleneck, not hardware. Brooks has consistently argued that the hard problem in robotics is not building the hardware but writing software that handles the infinite variety of real-world situations. The hardware progress in humanoid robots over the past three years has been remarkable. The software progress, while real, lags behind. Most deployed humanoid robots are operating with relatively narrow task libraries and require significant integration work for each new task.

The Rethink Robotics lesson

There is a deeper irony in Brooks’ position that deserves attention. Rethink Robotics, the company Brooks co-founded in 2008, failed in 2018 despite having many of the same ingredients that today’s humanoid companies claim will drive success. Rethink had a world-class team. It had significant funding. It had a clear market need (collaborative manufacturing robots). It had a visionary founder who understood robotics better than almost anyone alive.

It failed because the robots were not good enough. Baxter and Sawyer were too slow, too imprecise, and too limited in their task range to justify their cost for manufacturers who could hire human workers or buy traditional industrial robots. The technology was not ready.

Brooks learned from that failure, and it informs his current skepticism. He knows, from painful personal experience, that having a compelling vision, significant funding, and a clear market opportunity does not mean the technology is ready for prime time. The gap between “this should work” and “this does work, reliably, at scale, at a competitive price” swallowed $150 million and one of the most talented teams in robotics.

The counter-argument is that it is now eight years later, and the technology has improved dramatically. Large language models and vision-language-action models have transformed what robots can learn to do. Actuator costs have dropped. Sensor packages have gotten cheaper and more capable. The software infrastructure for training robot policies has advanced enormously. The conditions that killed Rethink Robotics may no longer apply.

But Brooks would likely respond: that is what they always say. The technology is always “dramatically better” than it was five years ago. The question is whether it is good enough. And “good enough” is a bar that robotics has consistently failed to clear on the timelines that advocates promise.

The seven deadly sins of prediction

In a widely read blog post titled “The Seven Deadly Sins of Predicting the Future of AI,” Brooks laid out a taxonomy of the mistakes people make when forecasting technology. These sins are worth reviewing because they apply directly to the current humanoid robot discourse.

Overestimating and underestimating. People simultaneously overestimate what a technology can do in the short term and underestimate what it can do in the long term. This is Amara’s Law, and Brooks has been one of its most effective communicators.

Imagining magic. Seeing a demo of a robot performing a task and concluding that the robot “understands” the task in the way a human would. Current humanoid robots that fold laundry have not developed a concept of “laundry.” They have learned a narrow policy for manipulating fabric objects that look like what they were trained on.

Performance versus competence. A robot that successfully sorts parts 95% of the time in a demo might fail 5% of the time in production. At industrial scale, a 5% failure rate is catastrophic. The distance from 95% to 99.99% is often longer than the distance from 0% to 95%.

Suitcase words. Words like “learning,” “understanding,” and “intelligence” that carry enormous amounts of meaning for humans but describe something much narrower when applied to current AI systems. When a company says their humanoid robot “learned” to perform a task, they mean something very different from what a human listener imagines.

Each of these sins is visible in today’s humanoid robot discourse. Every company video that shows a robot performing an impressive manipulation task invites the viewer to commit multiple prediction sins simultaneously.

The pattern-matching problem

Here is the fundamental tension. Brooks’ prediction methodology is based on pattern matching. He looks at a new technology claim, identifies which historical pattern it most closely resembles, and predicts that the outcome will be similar. This has worked brilliantly for 40 years because most technology hype cycles do follow similar patterns.

But pattern matching has a structural weakness: it cannot identify genuine phase transitions. There are moments in technology history when a new combination of factors creates a fundamentally different situation, one that does not follow previous patterns. The introduction of the iPhone in 2007. The emergence of deep learning after 2012. The launch of ChatGPT in 2022. In each case, a reasonable pattern-matching analysis would have predicted outcomes similar to previous cycles. In each case, the outcomes were radically different.

The question for humanoid robots is whether we are in a pattern-repeat or a phase-transition. Brooks’ methodology assumes pattern-repeat. The evidence could support either interpretation.

Three of the six key indicators look like a pattern-repeat of previous hype cycles. Three look like something genuinely new. That is not a clear verdict in either direction. It is a genuinely uncertain situation, which is exactly the kind of situation where both confident optimists and confident pessimists tend to be wrong.

What Brooks might be missing

If Brooks is wrong about this cycle, it will likely be for one or more of the following reasons.

The China factor has no historical analog. Brooks’ prediction framework was developed in the context of American and Japanese technology development, where the cycle from research to manufacturing scale-up is slow, capital-intensive, and constrained by high labor costs. China’s robotics ecosystem does not operate under these constraints. When Unitree can sell a humanoid robot for $16,000, they are operating in a manufacturing paradigm that Brooks’ historical pattern database may not account for.

Narrow deployment is enough to start. Brooks is right that general-purpose humanoid robots are years away. But the current deployment model is not general-purpose. It is narrow-task deployment in structured environments, essentially using a humanoid form factor to do work that is currently done by human hands because traditional automation cannot adapt to the physical workspace. This is less ambitious than what the hype promises, but it may be commercially viable now, not in 2035.

The venture capital dynamics are different. Previous robotics hype cycles were funded primarily by corporate R&D budgets that could be cut in a downturn. The current cycle is funded by venture capital, sovereign wealth funds, and public market investment at a scale that creates its own momentum. AgiBot raised $140 million in its first year. Figure AI raised over $750 million. This much capital can sustain companies through the “valley of death” between early deployment and profitability that killed Rethink Robotics.

Foundation models change the software equation. The software bottleneck that Brooks correctly identifies as the hardest problem in robotics may be addressed, at least partially, by the application of large foundation models to robot learning. Vision-language-action models are enabling robots to learn new tasks from demonstration far faster than was possible with traditional programming. This does not solve the generalization problem entirely, but it could compress the timeline from “decades” to “years.”

What Brooks is almost certainly right about

Even if the current humanoid cycle proves to be a genuine phase transition, Brooks is likely correct on several points.

The timelines promoted by most companies are too aggressive. Even the companies with the best track records are shipping hundreds of units per year, not the thousands or tens of thousands they are projecting. Manufacturing ramp takes time, even in China.

The demos overstate current capabilities. The gap between what humanoid robots can do in a controlled demo and what they can do reliably in production is still very large. Anyone making purchasing decisions based on demo videos is going to be disappointed.

General-purpose humanoid robots that can perform arbitrary household tasks are much further away than companies suggest. The factory deployment use case is tractable. The household use case is an order of magnitude harder.

The economics are still unproven at scale. Individual deployments may show positive ROI. But whether humanoid robots can compete with traditional automation and human labor across broad categories of work remains to be demonstrated.

The scorecard that matters

Brooks’ prediction methodology includes a feature that most technology pundits lack: accountability. He publishes his predictions with dates. He grades himself annually. He acknowledges when he was wrong. This alone puts him in the top 1% of technology commentators.

But accountability also means his predictions can be tested in real time. Here are the specific claims from his 2026 scorecard that we can track.

Will any single company ship more than 10,000 humanoid robots in a calendar year before 2028? Brooks would likely say no. Current trajectories from Unitree and AgiBot suggest it is possible but not certain.

Will a humanoid robot operate continuously in a factory environment for 8,000+ hours (roughly one year of single-shift operation) without major maintenance by 2028? This is Brooks’ reliability test. No public data suggests this has been achieved.

Will the total installed base of humanoid robots exceed 100,000 units by 2030? This requires roughly 7x growth per year from the current base. Ambitious but not mathematically impossible.

Will a humanoid robot successfully perform a novel household task (one it was not specifically trained for) on the first attempt, in an unmodified home, by 2030? This is the generalization test. Brooks almost certainly bets against it.

The reluctant conclusion

Rodney Brooks has earned his skepticism. He has been right more often than anyone else in this field, and right for the right reasons. His methodology is sound. His track record is extraordinary. He built real robots, sold real products, founded real companies, and learned from real failures. When he says that humanoid robot hype is running ahead of reality, the default assumption should be that he is correct.

And yet.

The current data contains features that genuinely do not match previous hype cycles. The shipment numbers are real and growing. The price points have crossed economically significant thresholds. The manufacturing ecosystem, particularly in China, operates at a speed and scale that Brooks’ historical framework may not fully account for. The software landscape has been transformed by foundation models in ways that were not foreseeable even three years ago.

The most honest answer is that we do not know yet whether Brooks is right about this cycle. We will know by 2028-2030. If global humanoid shipments reach 50,000-100,000 units by then, with demonstrated positive ROI in factory deployments, Brooks will have been too pessimistic for the first time in his career. If shipments plateau at 20,000-30,000 units and companies start folding or pivoting, Brooks will have called it correctly once again.

What we can say with confidence is this: whether Brooks is right or wrong about the timeline, his methodology remains the most valuable analytical framework in the field. Date-tethered predictions. Installation base analysis. The 3x rule for CEO optimism. The distinction between demos and deployments. The insistence on economic viability, not just technical possibility.

If the humanoid robot revolution does arrive on the timeline that optimists predict, it will not be because Brooks’ framework was wrong. It will be because the underlying conditions changed in ways that no historical pattern could have predicted. And if it does not arrive on that timeline, the people who dismissed Brooks will have to reckon with the fact that the most experienced roboticist alive was, once again, the most reliable forecaster.

Either way, anyone making decisions about humanoid robots, whether as an investor, a manufacturer, a policymaker, or an engineer, should be reading Rodney Brooks before they read anyone else. Not because he is always right. But because when he is wrong, he is wrong for specific, testable, well-documented reasons. That is more than almost anyone else in this field can say.