Brooks, Rodney

roboticist; former director of MIT CSAIL; co-founder of iRobot, Rethink Robotics, and Robust.ai

Rodney Brooks: Robotics

Lex Fridman Podcast · Sep 3, 2021

• Rodney Brooks
• robotics history
• Moravec paradox
• computation as metaphor
• Marvin Minsky and perceptrons
• reinforcement learning skepticism
• Tesla Autopilot and Waymo critique
• iRobot and Roomba
• Rethink Robotics and Baxter
• human-robot interaction
• Turing test critique
• MIT CSAIL culture
• AI hype cycles

Core ideas

Brooks traces his lifelong fascination with robotics to childhood in 1950s-60s Adelaide, building primitive learning circuits (an ice-cube-tray "computer" using copper-bridge electrolysis to learn simple associations) and unbeatable game-playing machines from telephone switchboard relays, framing this early drive as wanting to build something that could out-perform adults rather than any grand philosophical interest in intelligence. He is working on a book ("Not Even Wrong") arguing that "computation" — as formalized by Turing in 1936 to disprove one of Hilbert's problems — is a specific, historically contingent human metaphor (modeled on how human "computers" manually filled in mathematical tables) rather than a neutral, universal description of the physical world or the brain; he traces how this metaphor came to dominate not just AI but also neuroscience, artificial life, and even the discovery of DNA's structure, largely through figures connected to Warren McCulloch's circle (McCulloch and Pitts, von Neumann, Minsky), and warns that treating "computation" as if it explains consciousness or the universe risks the same category error as historical physical metaphors like "phlogiston" (illustrated by his example of a vibrating drum's stationary points, which the drum simply "has" without computing them).

Brooks endorses the Moravec paradox — that reasoning (chess, calculus) is comparatively easy for machines while perception and mobility are hard — explaining it as a consequence of evolutionary history: 600 million years were spent evolving perception and movement, while abstract reasoning, language, and culture are comparatively recent add-ons layered on top. He illustrates the depth of everyday perception/manipulation with an anecdote about his 16-month-old grandson spontaneously coordinating two different hand actions (pushing and turning a handle) to open an unfamiliar window, arguing this reflects an innate structural "pre-filter" that dramatically prunes the search space — something current reinforcement learning systems, which learn largely through undirected trial and error (he recounts visiting a secretive DeepMind lab where robot arms bashed around learning by pure exploration), notably lack. He is skeptical that scaling current techniques (citing GPT-3, AlphaZero, AlphaGo) constitutes progress toward general intelligence rather than impressive but narrow performance gains, tracing forms of self-play and reinforcement learning back to Donald Michie's 1961 tic-tac-toe program and Arthur Samuel's 1959 checkers program (built using surplus IBM mainframe computation time), and notes AlphaGo's total failure to generalize to a differently-sized board as evidence these systems lack the registration/grounding that human and animal perception achieves — while conceding he finds AlphaFold's protein structure predictions genuinely surprising, distinguishing "surprise" from "usefulness" (AlphaFold's accuracy is impressive but its lack of reliable confidence calibration limits practical use, though later versions have improved this).

On autonomous vehicles, Brooks argues public perception of progress is repeatedly distorted by people over-generalizing from demos (citing 1987 German highway autonomy tests and 1995 "hands-off" cross-country claims, both far earlier than commonly assumed), and is skeptical that vision-only systems like Tesla's will achieve full autonomy without both better camera dynamic range (cameras lack the human eye's roughly 11 orders of magnitude of range) and physical infrastructure changes — comparing this to how cars historically required society to redesign cities and laws (e.g., jaywalking statutes originally lobbied for by car manufacturers) and to how fully autonomous train systems required specially engineered environments (double platform doors, fully enclosed tracks). He nonetheless admits repeated surprise at how far vision-based engineering (Mobileye, then Tesla's from-scratch system under Andrej Karpathy's data engine) has advanced, and predicts near-term autonomous taxi deployment will be confined to constrained, low-speed environments (campuses, gated communities) rather than full city coverage, citing a personally observed incident where a Cruise vehicle froze for an extended period at a crosswalk, forcing a pedestrian with a stroller into the street — evidence, he argues, that current systems prioritize an overly conservative safety model that itself creates real-world risk and social friction.

Reflecting on his own companies, Brooks describes iRobot's proudest achievement as deploying battle-tested robots (with nine years of field experience defusing roadside bombs in Iraq and Afghanistan) to help shut down the Fukushima Daiichi plant in 2011, contrasting this with hyped-but-undeployed humanoid robots; he recounts the scrappy hardware-sourcing process behind the original Roomba (hunting for a sub-$1 chip with only 512 bytes of RAM in Hong Kong/Taiwan electronics markets to hit a $200 price target) and expresses regret that iRobot, despite ~19 years of Roomba sales, never found a comparably successful "next" home robot. At Rethink Robotics, he identifies getting safe, cage-free collaborative robot arms into factories as his proudest achievement, but describes in detail how he was gradually talked out of his original vision — a $3,000 plastic-geared, force-controlled arm aimed at customers with no existing robots — as customer expectations (demanding traditional 0.1mm positional repeatability rather than human-like force-guided precision) pushed the product toward metal gearboxes, dual arms, and a $25,000-35,000 price point that undermined the original strategy; the company ultimately failed after a Chinese-investor-linked acquisition fell through under CFIUS scrutiny, and its remaining team was absorbed by competitors within a week.

On AI hype and language, Brooks distinguishes his frustration with people who "over-extrapolate" from demos (predicting imminent solved-intelligence based on isolated benchmarks) from genuine entrepreneurial belief-driven risk-taking (comparing SpaceX's reusable rocket landing — built on decades of prior grid-fin and reusability research — favorably to the far less-grounded Hyperloop concept), while cautioning that Elon Musk's public "full self-driving" timeline claims blur this line in ways that affect stock markets and public expectations. On human-robot interaction, he argues robot "faces" and interfaces should never promise more capability than the system delivers (his design maxim: "if your robot looks like Albert Einstein, it should be as smart as Albert Einstein"), citing Baxter's simple animated eyes (used to telegraph the robot's next movement to nearby factory workers, not to simulate humanlike understanding) and mixed customer reactions to more anthropomorphic designs. He is skeptical of the Turing test as a meaningful intelligence benchmark (noting Turing himself intended it as a rhetorical device, not a serious test, and that it has degenerated into "gotcha" trivia games), and while open in principle to future AI-human emotional connection, believes we remain "a long, long way" from any AI achieving continuous, mutually evolving relationship-like engagement, since current systems lack persistent memory, goals, or any semblance of an ongoing "existence" between interactions.

He closes with reflections on MIT's AI Lab culture (recounting a story of accidentally installing a chip backward — causing visible smoke — while repairing a mainframe with Donald Knuth anxiously waiting to resume writing TeX overnight), a regret at never having asked figures like J.C.R. Licklider or Marvin Minsky (who ironically co-wrote the 1968 Perceptrons book that stalled neural network research despite his own PhD being on neural networks) more about their motivations before they died, skepticism that corporate AI labs (Google, Facebook/DeepMind) are meaningfully "leading" the field relative to universities despite vastly greater resources (arguing that locking away talented researchers under NDA, without public output, is often a net loss for the field), and advice to young people to avoid excessive conformity to prevailing academic incentives and be willing to risk repeated failure over a long career if aiming for genuine impact. On meaning, Brooks describes himself as an atheist who believes existence and complexity likely arose from chance rather than design, finds it near-certain (on emotional rather than evidentiary grounds) that intelligent life exists elsewhere in the universe, and reflects that being forgotten is the near-universal fate of essentially everyone who has ever lived — a fact he treats as sobering but ultimately compatible with finding deep, sustained joy in building robots.

Excerpts