This conversation is about roboticist Rodney Brooks' sobering take on AI and robotics. He argues most current AI 'breakthroughs' are just scaling up computation, not solving core intelligence—like how a 16-month-old can open a window with both hands on first try, while robots learn by crashing repeatedly. Brooks is cautious on self-driving cars, saying the real bottleneck is infrastructure and public acceptance, not tech. He highlights iRobot (Roomba) as a success story but notes no new home robot has been found in 19 years; Rethink Robotics (Baxter/Sawyer) failed because its price drifted from the target market; and Tesla's Autopilot is criticized for misleading naming and cameras that can't handle light changes as well as human eyes.
Rodney Brooks discussed the history and future of robotics on the Lex Fridman podcast. As a former director of MIT CSAIL and co-founder of iRobot, Rethink Robotics, and Robust.AI, he argued that the robotics industry should focus on learning "common sense" rather than over-relying on AI hype. Key po
Rodney Brooks — former director of MIT CSAIL, co-founder of iRobot/Rethink Robotics/Robust.AI, discusses the true boundaries of robotics technology with Lex Fridman. Core assessment: Most current "breakthroughs" in AI/robotics are engineering scale-ups rather than fundamental cognitive breakthroughs. True intelligence requires deep coupling between perception and action, precisely the problem that has remained unsolved for 60 years.
Brooks argues that the biggest problem with modern AI is treating "computation," a specific historical product, as the entirety of intelligence. He points out that the "computation" in Turing's 1936 paper essentially simulates the process of humans using pen and paper to perform mathematical operations—limited memory, simple rules, step-by-step execution. This model has been extremely successful in silicon-based implementations, but it does not represent the underlying logic of the universe.
"You hit a drum, the drumhead vibrates and creates nodes—the drum 'knows' where these points are, but it doesn't need to compute." In other words, physical systems can directly "know" certain things without symbolic operations. One of the core arguments of Brooks' new book Not Even Wrong is that humans tend to use familiar metaphors (computation, containers, positions) to understand everything, but these cognitive tools may be fundamentally inadequate in fields such as quantum mechanics and consciousness.
Four disciplines (neuroscience, AI, artificial life, and the origin of life) took shape simultaneously between 1945 and 1965, and all chose "computation" as their core metaphor. Brooks believes this is not because computation is the only correct framework, but because the founders of these fields at the time (McCulloch, von Neumann, Minsky, etc.) overlapped and influenced each other, forming an academic community centered on "computationalism."
Brooks argues that perception and movement are not "simple" problems; rather, they are core capabilities that evolution spent the most time refining. Over 600 million years, humans evolved perception and motor skills as multicellular organisms, while language, reasoning, and agriculture are merely "superstructures" layered on in the last tens of thousands of years.
He uses his 16-month-old grandson as an example: upon seeing a window for the first time, the child could push it with one hand while turning the handle with the other—coordinating two distinct actions to complete a task never encountered before. This is not random trial and error (reinforcement learning), but rather a kind of "pre-filtering" mechanism that dramatically narrows the search space. When DeepMind conducted RL experiments with a Sawyer robot, the machine could only learn through repeated collisions, a stark contrast to the cognitive efficiency of an infant.
"All the intelligence we know—humans, dogs, octopuses—operates by perceiving the world and taking action. The way they perceive objects is itself extraordinarily remarkable." Color constancy is one example: red is not a direct function of photons, but rather a construction of the brain based on the overall scene. Deep learning "solves" the labeling problem with large amounts of annotated data, but it does not solve the symbol grounding problem.
Brooks’ core critique of autonomous driving is not that the technology is infeasible, but that the underlying assumption—"keep all current roads, current vehicles, and current rules fixed, and simply replace the driver"—is fundamentally flawed. He looks back at history: after the automobile was introduced, cities were completely rebuilt and laws were rewritten (the term "jaywalking" was a crime promoted by the auto industry). The widespread adoption of autonomous driving will similarly require infrastructure support—dedicated lanes, enclosed tracks, access control systems, and so on.
"The U.S. has only 15–16 fully driverless rail transit systems, most of them at airports. The one in Honolulu was originally scheduled to open in 2017 and still hasn’t." In other words: even the most controllable environment—a rail system—has taken decades, and the difficulty of open roads is severely underestimated.
Brooks observes that a Cruise autonomous vehicle in San Francisco once stopped at a crosswalk because a pedestrian was opening a car door, blocking a baby stroller—a human driver would never have done that. This is not "safety," but a compromise on the safety of ordinary people who never consented to be part of the experiment. He believes the real breakthrough for autonomous driving will occur in limited domains (campuses, gated communities, low-speed environments), not city-wide deployment.
Regarding Tesla Autopilot, Brooks acknowledges its engineering achievements are impressive (especially the vision-only approach), but questions its insufficient dynamic range—cameras cannot handle 11 orders of magnitude in brightness variation like the human eye. He is more critical of the "Full Self-Driving" (FSD) name, which misleads consumers, and of Elon Musk’s excessive optimism about timelines.
Brooks initially set the goal for Rethink Robotics: a $3,000 force-controlled robot with plastic gearboxes, targeting new markets that had never used robots before. However, engineers complained that the "torque ripple" of the plastic gears made force control too difficult, and mechanical engineers promised to "fix it with a metal gearbox in six weeks" — which ended up taking two years. The CEO demanded a dual-arm design, driving the price up to $25,000. Customers (traditional factories) required 0.1mm repeat positioning accuracy rather than force control.
"I wanted to go to a market where no one had robots, but as the price rose, the only visitors were factories that already had robots — they judged the new product by old standards." In other words: once the product positioning deviates from the original target market, it falls into direct competition with mature products, and the differentiated advantages of the new technology (force control, safety, ease of programming) become disadvantages under old standards.
Ultimately, Rethink ran out of funds after CFIUS (the Committee on Foreign Investment in the United States) blocked an acquisition involving Chinese capital, and was sold off at 1/30 of its value. Brooks' reflection: "I let myself be persuaded to deviate from the original direction. I remember that meeting, I remember that day."
The success of Roomba was not a technological breakthrough, but an exercise in extreme cost engineering. Brooks recalls that to set the retail price at $199 (compared to the competitor Electrolux Trilobyte at €2,000), they had a computing budget of only 50 cents. He searched in Taipei, Hsinchu, and Hong Kong for "castrated" chips—6802 processors that had half their silicon removed but could still run—and eventually found a chip from Winbond with just 512 bytes of RAM.
"We had 14 failed business models before two winners emerged simultaneously in 2002. The board authorized production of 20,000 units, but we made 70,000—all sold out before Christmas." In other words: success is not linearly predictable; it is only achieved through extreme control over cost, market, and supply chain, after countless failures, and then by chance.
iRobot has been around for 19 years, yet its core product remains Roomba—"We’ve been looking for the next home robot, but we haven’t found it." This underscores the harsh reality of the consumer robotics market: after one hit, the next may never come.
Brooks argues that 2020 was not a "watershed year" for machine learning — it lacked the seismic impact of ImageNet in 2012. The number of papers surged, but "most are quite boring," representing incremental work rather than fundamental breakthroughs.
He expressed "surprise" at DeepMind's AlphaFold — the accuracy of protein structure prediction exceeded expectations — but immediately added: "You don't know which predictions are correct and which are wrong, which limits its practicality." In other words, even if the technology itself is impressive, its actual value remains questionable without interpretability and confidence assessment.
Regarding GPT-3, Brooks offered no direct evaluation, but the overall tone of the conversation implies: large-scale language models still operate within the "computational" framework by scaling up, rather than touching the essence of intelligence — perception, action, and continuous interaction with the world.
| Position | Analyst View | Key Data |
|---|---|---|
| iRobot (Roomba) | Bullish (success case) | Cumulative sales exceeded 30 million units; priced at $199 in 2002, computing budget 50 cents, chip with 512 bytes RAM |
| Rethink Robotics (Baxter/Sawyer) | Reflective (failure case) | Initial target $3,000 → final Baxter $25,000, Sawyer $35,000; burned $150 million in capital |
| Robust.AI | Not explicitly stated (current company) | Mission: teach robots "common sense" |
| Tesla Autopilot | Risk warning | Pure vision approach lacks dynamic range; FSD naming misleads consumers |
| Waymo/Cruise | Risk warning | Hundreds of driverless trips per week in Chandler, Arizona, but still with remote safety operators; Cruise in San Francisco blocked crosswalks due to pedestrians opening car doors |
| DeepMind (AlphaGo/AlphaFold) | Neutral (qualified recognition) | AlphaFold accuracy "surprising" but cannot determine which predictions are correct; AlphaGo would "completely fail" on a 21×21 board |
| SpaceX | Bullish (comparative case) | Vertical landing rockets had precedent (DC-X); grid fins existed in the 1960s; Elon Musk nearly went bankrupt but succeeded |
| Mercedes E450 | Bullish (personal use) | 2021 model, 360-degree surround view system "incredible" |
1. Brooks: Computation is not the fundamental logic of the universe; rather, it is a specific model of humans doing math with pen and paper, amplified by silicon. The drum "knows" the location of nodes without needing to compute—physical systems can "know" directly.
2. Brooks: Moravec's paradox is not a paradox; it reflects the timescale of evolution. 600 million years honed perception and movement, while a few tens of thousands of years layered on language and reasoning—we have reversed our understanding of "hard" and "easy."
3. Brooks: A 16-month-old infant can coordinate both hands to open a window they have never seen before, while DeepMind's Sawyer can only learn through repeated collisions. Humans have some "pre-filtering" mechanism that dramatically narrows the search space; RL does not.
4. Brooks: The lesson from Rethink Robotics—low price does not equal low-end, but once the price rose to $25,000, the only visitors were factories that already had robots, judging the new product by old standards. Misalignment of product positioning with the target market is a fatal mistake.
5. Brooks: The real bottleneck for autonomous driving is not technology, but infrastructure and public acceptance. The U.S. has only 15-16 fully unmanned rail transit systems, most at airports—even rail took decades.
6. Brooks: iRobot's success came from extreme cost control after 14 failures—a $0.50 computing budget, 512 bytes of RAM, and selling out 70,000 units before Christmas. The next home robot has not been found in 19 years.
7. Brooks: The current explosion in AI papers is "basically boring"; 2020 was not a watershed year. AlphaFold is surprising, but "you don't know which predictions are correct."
8. Brooks: The criticism of Elon Musk is not about questioning his conviction, but that he fails to distinguish between "technology is mature" and "still far off." SpaceX had precedents to follow; Hyperloop did not—but he talks about them the same way.