Moravec's paradox is the observation that tasks requiring abstract, high-level reasoning, playing a board game, proving a theorem, filling out a tax form, take relatively little computation for a machine, while the sensorimotor and perceptual skills that come naturally to a healthy one-year-old, recognizing a face, picking up an oddly shaped toy, crossing a cluttered room without falling, demand enormous computational resources and remain difficult for machines to reproduce reliably. The pattern was described independently by several robotics and AI researchers between the mid-1970s and late 1980s. It reached its standard formulation in Carnegie Mellon University roboticist Hans Moravec's 1988 book Mind Children [1], and is also associated with roboticist Rodney Brooks [2] and AI researcher Marvin Minsky [2][9], who reached related conclusions independently. Linguist Steven Pinker later popularized the idea outside robotics circles, writing that "the main lesson of thirty-five years of AI research is that the hard problems are easy and the easy problems are hard" [10]. Nearly four decades later, the same asymmetry sits at the center of humanoid robot development: it is easier to build a system that holds a fluent conversation or solves competition-level mathematics than one that reliably folds laundry or spreads peanut butter on bread [11][14].

In brief

Picture a state-of-the-art AI system that can write software, solve calculus problems, or beat a grandmaster at chess. Now picture a robot in the same room, wired to the same computer, trying to pick a dropped pen up off a carpeted floor. The AI handles the "hard" intellectual task easily; the robot may well fail at the "easy" physical one. Moravec's paradox names that mismatch. It exists, on the standard explanation, because reasoning is a young and comparatively simple skill that humans perform consciously and slowly, while perception and movement are ancient, heavily optimized skills that run automatically, outside conscious awareness, which makes them far harder to study and copy into a machine [1][9].

Origins: reasoning is cheap, perception is not

Moravec, who joined Carnegie Mellon University's Robotics Institute in 1980 after a PhD at Stanford under John McCarthy, had been circling the idea since the late 1970s [3]. In a 1977 essay he argued that "the most difficult tasks to automate, for which computer performance to date has been most disappointing, are those that humans do most naturally, such as seeing, hearing and common sense reasoning," attributing the gap to raw processing power: computers of the era, he wrote, were "still a hundred thousand to a million times too slow to match the performance of human nervous systems" in perception and motor control [4]. Checkers and simple logical proofs, though effortful for a person, could already be handled with far less computation, because they involve a narrow, well-defined state space and little raw sensory data [4].

Moravec gave the idea its lasting form a decade later in Mind Children: "It is comparatively easy to make computers exhibit adult level performance on intelligence tests or playing checkers, and difficult or impossible to give them the skills of a one-year-old when it comes to perception and mobility" [1]. The choice of checkers rather than chess is a period detail worth keeping: in 1988, checkers programs were already a mature success story, while chess software would not defeat a reigning world champion until IBM's Deep Blue beat Garry Kasparov in 1997. Later commentary often substitutes chess or Go, since both have since fallen to machines, but the underlying claim is unchanged: well-defined symbolic games are cheap for computers, embodied competence is not [2].

The evolutionary and reverse-engineering explanation

Moravec's explanation was evolutionary. He argued that sensorimotor and perceptual skills are not actually simple; they only feel simple because they are extremely old and extremely well optimized by natural selection. "Encoded in the large, highly evolved sensory and motor portions of the human brain is a billion years of experience about the nature of the world and how to survive in it," he wrote. "The deliberate process we call reasoning is, I believe, the thinnest veneer of human thought, effective only because it is supported by this much older and much more powerful, though usually unconscious, sensorimotor knowledge" [1]. On this view, walking, catching a thrown object, and recognizing a face are the products of hundreds of millions of years of selection, refined across an unbroken line of ancestors for whom failure at these tasks meant death before reproducing. Abstract reasoning, arithmetic, and formal logic are comparatively new, plausibly under 100,000 years old, and have not been under nearly as much selective pressure to become efficient [1][2].

This has a second, practical consequence beyond raw computational demand: it makes sensorimotor skill much harder to reverse-engineer. A skill performed consciously and step by step, such as long division, can be inspected and translated into a program relatively directly, because the person doing it has some introspective access to the steps. A skill that runs automatically and unconsciously, such as adjusting grip pressure on an egg without looking at it, cannot be interrogated the same way; the person doing it effortlessly usually cannot explain how. AI researcher Marvin Minsky made close to the same point in his 1986 book The Society of Mind: "In general, we're least aware of what our minds do best," and "we're more aware of simple processes that don't work well than of complex ones that work flawlessly" [9]. The tasks that feel hard to a person are hard partly because the effort of doing them is conscious; the tasks that feel easy are easy specifically because the hard work behind them is hidden from awareness, which is exactly what makes that work resistant to being copied into code.

Brooks and embodied intelligence

Rodney Brooks, an Australian-born roboticist who joined the MIT faculty in 1984 and later directed MIT's Artificial Intelligence Laboratory and its successor, CSAIL, from 1997 to 2007, reached a related conclusion from a different direction [8]. Rather than starting from the mismatch between checkers and grasping, Brooks started from the failure of classical AI, which tried to give robots an internal symbolic world model to reason over before acting, to produce robots that could reliably move through a real room. In papers including "A Robust Layered Control System for a Mobile Robot" (1986) and "Elephants Don't Play Chess" (1990), Brooks argued that intelligence in animals, and by extension in robots, is not built on abstract representation and planning, but on tight, largely reactive loops between sensing and acting, layered on top of one another [5][7]. His title made the point directly: an elephant has no capacity for chess, and nobody would call it unintelligent for that reason. Brooks's "subsumption architecture," and the broader approach he called "Nouvelle AI," dispensed with detailed internal world models in favor of simple behaviors wired close to sensors and motors, on the premise that "the world is its own best model" and does not need to be re-represented inside a machine to act on it effectively [6][7]. Where Moravec explained the paradox through computation and evolutionary time, Brooks treated it as evidence that classical AI had been solving an easier, disembodied version of intelligence and mistaking it for the whole problem.

Pinker's summary: the hard problems are easy

The idea reached a much wider audience through Steven Pinker's 1994 book The Language Instinct, written partly to argue that language is a specialized, evolved capacity rather than a general byproduct of intelligence. Pinker used Moravec's paradox as his central illustration of how misleading human intuitions about difficulty can be: "The main lesson of thirty-five years of AI research is that the hard problems are easy and the easy problems are hard," he wrote, adding that "the mental abilities of a four-year-old that we take for granted, recognizing a face, lifting a pencil, walking across a room, answering a question, in fact solve some of the hardest engineering problems ever conceived" [10]. Pinker's framing helped establish the paradox as a standard reference point well outside robotics, cited across linguistics, cognitive science, and popular discussion of automation ever since [2][10].

Illustrative examples of the reversal

Task type	Example	Status for AI or robots as of 2026
Symbolic games	Checkers, chess, Go	Superhuman since Deep Blue (1997) and AlphaGo (2016)
Formal mathematics	Olympiad-level problem sets	Frontier language models reached gold-medal-equivalent scores at the 2025 International Mathematical Olympiad [22]
Closed-book knowledge tasks	Standardized exams, coding tests	At or above top-human percentile for leading models
Basic mobility	Walking on flat or lightly cluttered ground	Reliable in many commercial humanoids and quadrupeds
Dynamic locomotion	Backflips, parkour-style movement	Demonstrated by humanoid and quadruped robots, trained largely in simulation [17]
Fine manipulation	Folding a shirt, spreading peanut butter, turning a key	Only recently and inconsistently solved; an active research frontier [11][14]
Everyday perception in clutter	Finding and grasping an unfamiliar object on a messy counter	Still a leading cause of failure in real-world robot deployments [14][19]

Why perception and robot manipulation resist computation

The evolutionary story explains why people underestimate the difficulty of sensorimotor tasks, but modern computer science adds a more concrete, engineering-level account of why perception and robot manipulation are so expensive to automate.

Reasoning tasks such as chess, theorem proving, or arithmetic operate over small, cleanly defined state spaces: a finite board, a fixed rule set, complete information about the current position. Perception and manipulation operate over continuous, high-dimensional, noisy, and only partially observable environments. A camera returns millions of pixels that must become an estimate of an object's shape, material, and pose before a robot can decide how to grasp it, and that estimate is never perfect [2]. Computer vision research spent decades on that translation step before deep neural networks made it tractable at scale, and recognizing an object is still only the first of several hard problems; predicting how it will behave once touched, whether it slips, deforms, or tips, is a separate and harder one.

Simulation makes the gap concrete. Locomotion and dynamic movement, walking, running, even backflips, can largely be learned in physics simulators and transferred to real hardware, because rigid-body dynamics and ground contact are relatively well modeled computationally, and humanoid and quadruped robots from companies such as Boston Dynamics and Unitree have performed acrobatics trained mostly in simulation [17]. Manipulation is a harder simulation problem: it requires modeling friction, deformable materials, and fine contact forces accurately enough that a policy learned in simulation still works on a real, unpredictable object, a gap researchers try to close with sim-to-real transfer techniques [17]. A 2025 analysis of the trend put it this way: training a robot to backflip is closer to training a blind gymnast, since balance and joint trajectories can be worked out almost entirely offline, while teaching the same robot to see and manipulate an unfamiliar object demands contact-rich, visually grounded data that simulation still cannot supply [17].

There is also a data asymmetry. Large language models learn to reason partly because the internet contains an enormous amount of text that implicitly teaches reasoning: code, mathematics, argument, explanation. There is no equivalent public archive of how to move an arm to clean a greasy pan; physical competence mostly is not written down anywhere, because, following Minsky's and Moravec's observations, the people who have it cannot fully articulate it [9][11]. Robotics has responded by trying to build that missing archive directly, an effort sometimes grouped under robot learning: through imitation learning from human teleoperation and video, reinforcement learning in simulation and on real hardware, and increasingly large vision-language-action models and other robot foundation models trained on cross-robot demonstration data [11][17]. Low-cost tools such as the handheld universal manipulation interface exist specifically to gather that kind of embodied training data cheaper and faster, on the premise that Moravec's paradox, today, is substantially a data problem rather than a purely algorithmic one [11][21].

The robot olympics and the manipulation bottleneck

The clearest recent demonstration of Moravec's paradox has come from humanoid robotics rather than academic writing. In late 2025, Benjie Holson, an engineer at the robotics company Robust.AI (cofounded by Rodney Brooks), proposed household benchmark tasks he called the "Humanoid Olympic Games," organized into five categories, doors, laundry, tools, fingertip manipulation, and wet manipulation, each with bronze, silver, and gold tiers [12][13]. The challenges were deliberately mundane: turning a key in a lock, spreading peanut butter, washing a greasy pan, rolling a sock right side out, hanging a buttoned dress shirt. Holson's stated motivation was that flashier competitions, including a state-run event pitting humanoid robots against each other in sports-style contests, missed what people actually want a home robot to do, which is chores, not martial arts [15].

Holson initially expected the hardest tiers to take years to solve [14]. Instead, the robotics company Physical Intelligence, applying its pi0.6 vision-language-action model, completed 11 of the 15 challenges, from bronze through gold, within about three months, publishing its results in a December 2025 post titled "Moravec's Paradox and the Robot Olympics" [11]. The company reported gold-level results on three of five task categories and silver on the other two, with an overall task success rate of 52 percent and average task progress of 72 percent, most of it achieved with under nine hours of task-specific data collection per skill. A general-purpose vision-language model without robotics-specific pretraining, tested as a baseline, completed none of the tasks and reached only 9 percent average progress [11]. These systems relied on ordinary cameras rather than dedicated force or tactile sensors, inferring contact and grip forces visually, a detail independently confirmed in Scientific American's coverage of the same results [14]. Physical Intelligence summarized the issue in Moravec-like terms: people cannot program physical intelligence into a robot because they do not fully understand it at a conscious level, so the practical route is to learn it from diverse, real physical interaction data instead [11].

The results reframed the timeline more than the underlying difficulty. Holson revised his own estimate for reliable, commercially deployable home robots from roughly fifteen years down to about six, while cautioning that the remaining gap is mostly about reliability, generalization to new objects and environments, and safety, not raw task feasibility [14]. IEEE Spectrum and Scientific American both framed the episode as a real-time illustration of Moravec's paradox: a robot landing a backflip, largely the product of simulation-trained reinforcement learning, had already become almost routine, while one reliably folding a shirt or peeling an orange remained newsworthy precisely because it was still hard [13][14][17].

The episode also revived a narrower debate about robot hardware: whether closing the manipulation gap requires human-like, fully actuated, five-fingered dexterous hands, or whether simpler end effectors are enough once perception and control software improve. A 2025 survey of robotic hand designs found that complex anthropomorphic hands are not necessary for most manipulation tasks. In the DARPA Robotics Challenge, for example, 15 of 25 competing teams used simpler three- or four-fingered underactuated hands, and none used a fully actuated humanlike hand; the authors concluded that wrist mobility and finger abduction matter more for practical dexterity than raw finger count or total degrees of freedom [16]. That finding sits inside Moravec's paradox rather than against it: mimicking human anatomy is not automatically the fastest way to close the manipulation gap, and tactile sensing can sometimes be substituted with vision alone [11][16].

The framing has also reached the industry's largest supplier of AI hardware. At CES 2026, Nvidia chief executive Jensen Huang described a "ChatGPT moment for physical AI," saying machines are now beginning to "understand, reason and act in the real world," and positioned physical AI and embodied AI for humanoid robots as the next major computing shift after language models [18]. Whether that shift arrives on Huang's optimistic timeline, or the slower one implied by Moravec's paradox, remained, as of mid-2026, an open question rather than a settled one [19][20].

Reassessment and criticism

Not every AI researcher accepts that Moravec's paradox, as usually stated, is a well-established empirical finding rather than a memorable slogan. In a January 2026 essay, Princeton computer scientist Arvind Narayanan argued that the paradox "has never been empirically tested" in any rigorous way, despite decades of repetition by respected figures [19]. His central objection is a selection-bias argument: commentary on AI tends to highlight tasks that are hard for humans and easy for machines, or the reverse, while ignoring tasks that are easy for both and tasks that are hard for both, which can create the appearance of a pattern that is not really there [19]. Narayanan also challenged Moravec's evolutionary explanation directly, arguing that abstract reasoning may depend on older sensorimotor and common-sense machinery rather than being independent of it, which would help explain why AI systems that are superhuman at chess or narrow mathematics still struggle with open-ended reasoning in fields such as law or science, where, unlike chess, there is no complete rule set and no fast, unambiguous feedback signal [19]. His conclusion is that the paradox, whatever its rhetorical value, is not a reliable predictor of which tasks AI will master next [19].

An earlier, more measured critique came from AI researcher and writer Nathan Lambert, who argued in 2023 that the paradox rests on empirical observation rather than deep theory, and that robotics faces a compounding problem: unlike text, embodied training data cannot be scraped from the internet, it has to be generated through real or simulated interaction [20].

That critique narrows the pattern more than it erases it. Even skeptics of the strong version of Moravec's paradox generally accept the narrower claim underneath it: tasks with clean formal structure and complete information, board games, symbolic mathematics, closed-book exams, have consistently proven easier to automate at a given stage of AI development than open-ended physical tasks in an unstructured environment. Humanoid robotics in 2026 is still organized substantially around closing that gap [11][14][19][20].

See also

References

Hans Moravec, Mind Children: The Future of Robot and Human Intelligence, Harvard University Press, 1988.
"Moravec's paradox," Wikipedia, accessed July 2026. https://en.wikipedia.org/wiki/Moravec%27s_paradox
"Hans Moravec," Wikipedia, accessed July 2026. https://en.wikipedia.org/wiki/Hans_Moravec
Hans Moravec, "Intelligent Machines: How to Get There From Here and What to Do Afterwards," 1977, hosted by the Carnegie Mellon University Robotics Institute. https://frc.ri.cmu.edu/~hpm/project.archive/general.articles/1977/smart
Rodney A. Brooks, "A Robust Layered Control System for a Mobile Robot," IEEE Journal of Robotics and Automation, 1986.
Rodney A. Brooks, "Intelligence Without Representation," Artificial Intelligence, vol. 47, 1991. https://people.csail.mit.edu/brooks/papers/representation.pdf
Rodney A. Brooks, "Elephants Don't Play Chess," Robotics and Autonomous Systems, vol. 6, 1990.
"Rodney Brooks," Wikipedia, accessed July 2026. https://en.wikipedia.org/wiki/Rodney_Brooks
Marvin Minsky, The Society of Mind, Simon and Schuster, 1986.
Steven Pinker, The Language Instinct: How the Mind Creates Language, William Morrow, 1994 (Perennial Modern Classics edition, HarperCollins, 2007).
Physical Intelligence, "Moravec's Paradox and the Robot Olympics," December 22, 2025. https://www.pi.website/blog/olympics
Benjie Holson, "Benjie's Humanoid Olympic Games," General Robots (Substack), 2025. https://generalrobots.substack.com/p/benjies-humanoid-olympic-games
IEEE Spectrum, "Humanoid Robot Olympics: Tackling Everyday Chores," November 4, 2025. https://spectrum.ieee.org/humanoid-robot-olympics
Deni Ellis Bechard, "Why Humanoid Robots Are Learning Everyday Tasks Faster Than Expected," Scientific American, published online March 2, 2026. https://www.scientificamerican.com/article/why-humanoid-robots-are-learning-everyday-tasks-faster-than-expected/
John Koetsier, "The Robot Olympics Will Have Zero Sports. Here's Why," Forbes, December 24, 2025. https://www.forbes.com/sites/johnkoetsier/2025/12/24/the-robot-olympics-will-have-zero-sports-heres-why/
Alexander Fabisch, Wadhah Zai El Amri, Chandandeep Singh, and Nicolas Navarro-Guerrero, "Do Robots Really Need Anthropomorphic Hands? A Comparison of Human and Robotic Hands," arXiv:2508.05415, submitted August 2025, revised 2026. https://arxiv.org/abs/2508.05415
"Backflipping Robots and Moravec's Paradox," Hardware FYI (Substack), July 29, 2025. https://hardwarefyi.substack.com/p/backflipping-robots-and-moravecs
Axios, "Nvidia CES 2026: Jensen Huang says 'ChatGPT moment for physical AI' is coming," January 5, 2026. https://www.axios.com/2026/01/05/nvidia-ces-2026-jensen-huang-speech-ai
Arvind Narayanan, "Fact checking Moravec's paradox," normaltech.ai, January 29, 2026. https://www.normaltech.ai/p/fact-checking-moravecs-paradox
Nathan Lambert, "Can robotics take off like GenAI? Moravec's paradox vs. scaling laws," Interconnects, September 29, 2023. https://www.interconnects.ai/p/robotics-vs-moravecs
Cheng Chi et al., "Universal Manipulation Interface: In-The-Wild Robot Teaching Without In-The-Wild Robots," arXiv:2402.10329, 2024. https://umi-gripper.github.io/
Google DeepMind, "Advanced version of Gemini with Deep Think officially achieves gold-medal standard at the International Mathematical Olympiad," 2025. https://deepmind.google/blog/advanced-version-of-gemini-with-deep-think-officially-achieves-gold-medal-standard-at-the-international-mathematical-olympiad/

Editing: Moravec's paradox