Lex Fridman PodcastRodney Brooks: Robotics | Lex Fridman Podcast #217
CHAPTERS
- 0:00 – 4:53
DOMO the humanoid torso: beauty, actuation, and not overpromising with robot faces
Rodney Brooks names DOMO as the most beautiful robot he’s worked with, describing its mechanical artistry and series elastic actuation. The discussion expands into how a robot’s appearance sets expectations, and why faces can "overpromise" intelligence.
- •DOMO’s design: upper-torso humanoid with actuated eyeballs/cameras and three-finger hands
- •Series elastic actuators, cable drives, and the robot as mechanical art
- •Robot appearance as a promise: don’t imply capabilities you can’t deliver
- •"If it looks like Einstein, it should be as smart as Einstein"
- •Using eyes/eye-like displays to communicate intent (Baxter/Sawyer preview)
- 4:53 – 7:15
First robots and first ‘brains’: Wonder Books, circuits, and an ice-cube-tray learning machine
Brooks traces his childhood fascination with robots back to early books on electricity and ‘giant brains.’ Limited by mechanics and budget, he focused on building learning and logic systems—famously including a chemical learning setup using an ice cube tray.
- •Early inspiration: How and Why Wonder Books on electricity and robots
- •Childhood attempts to build robot ‘brains’ more than robot bodies
- •Chemical learning network using copper bridges in an ice cube tray
- •Progression to transistors, logic gates, flip-flops
- •Early joy: building systems that outperform adults in games
- 7:15 – 8:35
Can machines think? Humans as machines—and the harder question of how to build minds
Prompted by Turing’s ‘Can machines think?’, Brooks argues that humans are machines and therefore machines can think in principle. But he emphasizes that feasibility—whether we’re smart enough to build such systems—is the real unknown.
- •Brooks’ stance: humans are machines; thinking is a physical process
- •Separating ‘possible in principle’ from ‘we know how to do it’
- •Skepticism about our ability to engineer human-like minds soon
- •Mind may be distributed beyond the brain (body, gut nervous system)
- •Social interaction and externalized knowledge as part of intelligence
- 8:35 – 25:36
‘Not Even One Wrong’: computation’s history and why the brain-as-computer metaphor misleads
Brooks outlines his book-in-progress arguing that our dominant notion of computation is historically contingent and metaphor-laden. He revisits Turing’s original framing and tracks how computation became a catch-all explanatory tool for mind and life.
- •Computation’s lineage: Napier/Kepler → Babbage/Lovelace → Turing (1936)
- •Turing machine as a model of human pencil-and-paper procedure
- •‘Computation’ as a chosen definition, not a universal property like charge
- •The 2x2 matrix: AI vs artificial life; neuroscience vs abiogenesis
- •Human metaphors (place/container) shaping how we reason about intelligence
- 25:36 – 33:56
Perception, grounding, and Moravec’s ‘paradox’: why robots struggle with the real world
Brooks argues that intelligence we recognize is tightly coupled to perception and action in the world. He critiques deep learning’s successes as insufficient for grounding and robust understanding, stressing that perception and registration remain hard.
- •All known intelligences perceive and act; real-world perception is complex
- •Color constancy and context: why ‘red stop sign’ isn’t trivial to infer
- •Labeling ≠ grounding: the symbol grounding problem remains
- •Registration as a deep philosophical/engineering challenge (Brian Cantwell Smith)
- •Moravec’s paradox reframed via evolution’s long investment in sensorimotor skills
- 33:56 – 40:18
Manipulation and learning: a toddler opens a window, DeepMind’s Sawyer robots, and pruning the search space
A story about Brooks’ grandson reveals the sophistication of human manipulation and generalization. Brooks contrasts children’s targeted exploration with brute-force reinforcement learning and describes visiting DeepMind’s secret robot experiments using Sawyer arms.
- •16-month-old coordinating two hands to open an unfamiliar window
- •Kids explore mechanisms with strong priors; not random search
- •DeepMind visit: reinforcement learning on Sawyer robots (and the secrecy irony)
- •RL can work, but the challenge is building the pruning/priors humans use
- •Skepticism toward recurring ‘now we’ve got it’ moments in AI history
- 40:18 – 55:36
What deep learning genuinely surprised him: ImageNet, AlphaFold, and the many meanings of ‘learning’
Lex presses Brooks on whether any recent AI breakthroughs impressed him; Brooks concedes ImageNet performance was shocking and picks AlphaFold as the most surprising. He also warns that words like ‘learning’ and ‘smart’ are suitcase terms that hide many distinct phenomena.
- •ImageNet as a legitimate ‘shock’ moment, but not full robot vision
- •AlphaFold as the standout surprise—paired with concerns about knowing when it’s right
- •Game AIs: impressive but narrow; generalization failures (e.g., different Go board sizes)
- •‘Learning’ as a suitcase word: learning chess vs biking vs navigation
- •Hype cycles: scaling up techniques can fool people about general intelligence
- 55:36 – 1:16:25
Self-driving cars: impressive driver-assist tech, long-tail edge cases, and infrastructure as the missing piece
Brooks praises modern driver-assist features while arguing the industry keeps replaying a familiar overgeneralization story from early demos. He predicts public tolerance for AV-caused deaths will be far lower than for human-caused deaths, and expects progress to accelerate with infrastructure changes and constrained domains.
- •1987 autonomous freeway demo and 1995 CMU cross-country as cautionary historical anchors
- •Why AV risk tolerance is asymmetric vs human driving deaths
- •Cars vs trains: autonomy succeeds where environments are engineered (platform doors, fenced tracks)
- •Dedicated lanes and modified road markings as practical enablers
- •SF anecdotes illustrate ‘stuck’ behaviors and externalized safety costs to bystanders
- 1:16:25 – 1:26:44
Optimism vs overpromising: believing in the impossible without misleading users or markets
Lex argues big belief is required to achieve breakthroughs; Brooks agrees but draws a line where overpromising demoralizes teams and misleads customers. They discuss Elon Musk’s strengths, where he’s ‘wrong sometimes,’ and the downstream effects of ambitious claims on trust and investment.
- •Historical leaps: Wright brothers to moon landing within one lifetime
- •Brooks’ frustration: people leap from demos to ‘we’re basically done’
- •SpaceX praised; Hyperloop framed as far less mature technologically
- •Naming and marketing matter: ‘Full Self-Driving’ as a promise people pay for
- •Ambitious deadlines can motivate—but can also distort markets and expectations
- 1:26:44 – 1:38:02
Prediction accountability: Brooks’ 2018–2050 forecast list and what it reveals about progress
Brooks explains why he published dated predictions and commits to yearly scorekeeping through 2050. He revisits AV timelines, admits optimism in some entries, and argues real deployments will start in bounded, low-speed environments and only later expand.
- •Predictions started Jan 1, 2018; annual reviews planned for 32 years
- •Accountability as an antidote to forgotten failed forecasts
- •Cambridgeport and SF as examples of hard, messy urban driving domains
- •Driverless services first in campuses/gated communities/retirement villages
- •Trolley problem dismissed as misframing: real driving is ‘brake hard,’ not moral arithmetic
- 1:38:02 – 1:48:01
iRobot’s real-world impact: war-zone deployments, Fukushima, and the Roomba cost-engineering grind
Brooks highlights iRobot’s field experience as the reason its robots could help at Fukushima, contrasting it with ‘showpiece’ robotics. He then details the Roomba’s ruthless cost constraints—down to selecting microcontrollers with hundreds of bytes of RAM—and the entrepreneurial reality of repeated failed business models before a hit.
- •Fukushima response enabled by prior deployment of ~6,500 robots in Iraq/Afghanistan
- •Critique of hype robotics (e.g., humanoids that walk but aren’t deployable)
- •Roomba target: consumer price point ($200 shelf price) driving extreme engineering choices
- •Sourcing cheap computation: trips to Taiwan/Hong Kong; 50-cent compute budget
- •iRobot’s path: 14 failed business models before the 2002 breakout
- 1:48:01 – 1:57:18
Rethink Robotics: collaborative arms, the $3k dream, and how product-market fit drifted
Brooks describes Rethink’s core achievement—safe, codelessly retaskable robots outside cages—while recounting painful pivots that raised cost and shifted the target customer. He explains force control vs position repeatability and how factory expectations pushed the product into a harder market with mismatched requirements.
- •Pride point: safe cobots and re-tasking without writing code
- •Original vision: $3,000 force-feedback robots for non-robot factories
- •Plastic gearbox prototypes vs control challenges (torque ripple)
- •Cost/feature creep: Baxter at ~$25k, later $35k arm; drift toward incumbent-factory expectations
- •Company outcome influenced by acquisition dynamics and CFIUS/financing realities
- 1:57:18 – 2:05:28
Human connection, love, and the Turing test: continuity, intention, and why ‘fooling’ is the wrong goal
The conversation turns to emotional bonds with AI, with Brooks predicting we’re far from mutual romantic partnership but acknowledging humans already form attachments to machines. He critiques the Turing test’s evolution into trickery and argues that meaningful companionship requires continuity of memory and apparent goals or intentions.
- •Romantic AI: humans may ‘fall in love’ easily; reciprocal love is far away
- •Sci-fi fallacy: holding the world constant while changing one technology
- •Bicentennial Man as a better model: co-evolution of humans and machines
- •Why voice assistants feel shallow: no shared history, continuity, or persistent goals
- •Turing test as rhetorical device vs modern ‘gotcha’ game (AI Olympics story)
- 2:05:28 – 2:17:05
MIT/Stanford eras and AI legends: regrets, Minsky questions, and early glimpses of the future
Brooks reflects on working alongside AI pioneers and regrets not interviewing them when he had the chance. He shares the thrill of early computing infrastructure (video terminals, printers, Lisp machines) and recounts a Don Knuth late-night mainframe repair story that ended in literal smoke.
- •Regret: didn’t ask foundational figures (e.g., Licklider) the questions he now needs
- •Minsky’s Perceptrons impact: why attack neural nets so hard?
- •Corporate labs ‘locking away’ talent vs university openness
- •Early tech wonder: email since 1977; Lisp machines as proto-PCs
- •Don Knuth story: fixing a disk system, inserting a chip backward, smoke, then recovery
- 2:17:05 – 2:24:51
Advice, mortality, and meaning: taking unsafe bets, finishing the book, and order from disorder
Brooks advises young people to resist herd pressures, accept frequent failure, and make ‘unsafe’ choices if they want real impact. The discussion ends with reflections on aging, legacy, atheism, and the mystery of how complexity and order emerge in a seemingly random universe.
- •Career advice: impact requires risk-taking and repeated failure over a long horizon
- •Fear of decline more than death itself; COVID sharpened urgency
- •Legacy hope: finish the book and change one reader’s thinking
- •Meaning of life: humans fixate on the immediate; big-picture thinking is hard
- •Awe at complexity: pockets of order, potential alien life, and inevitable forgetting
