Robotics conversations tend to split into two boring camps: engineers talking specs nobody outside the lab cares about, or futurists talking about robot uprisings that aren't coming. The episodes below, pulled from our full library of summaries, mostly skip both traps. They're built from long-form interviews with the people who actually shipped the Roomba, built Waymo's lidar stack, invented the motion-planning algorithm your self-driving car quietly relies on, or spent a decade making 'shitty robots' before building a real company.
Expect specific numbers, not vague optimism: what a Cozmo robot actually cost to make, how many miles Waymo simulates for every mile it drives, why a $3,000 robot design got killed in favor of a $30,000 one. Whether you're into home robots, autonomous vehicles, drones, or the reinforcement-learning theory underneath all of it, there's a full breakdown waiting below.
Boris Sofman: Waymo, Cozmo, Self-Driving Cars, and the Future of Robotics | Lex Fridman Podcast #241
Sofman ran two very different robotics chapters: he built Cozmo, the cheap four-degree-of-freedom robot that proved character beats humanoid form, and now leads Waymo's autonomous trucking effort. The Cozmo section is a gut punch, Anki hit close to $100M in revenue with 85 percent of it crammed into Q4, and cash-flow physics killed the company anyway. On the Waymo side, he argues the hardest problem in self-driving isn't the autonomy, it's the evaluation, since a truck driver only sees a serious freeway event every 1.3 million miles. Listen if you want the real story of why brilliant robotics companies still go bankrupt, and how you actually validate a system for events that rare.
Read the full episode notesSimone Giertz: Queen of Sh*tty Robots, Innovative Engineering, and Design | Lex Fridman Podcast #372
Giertz built her reputation on deliberately useless machines, then pivoted to running a real product company, all while being diagnosed with a golf-ball-sized brain tumor the day she walked off her TED Talk stage. She's candid about the self-deprecating 'idiot' persona being a defense mechanism for surviving online as a woman in engineering, and about refusing second opinions on her surgery because uncertainty was too terrifying to sit with. This one's less about robot specs and more about the person behind them, worth it for anyone who wants the human cost of building in public.
Read the full episode notesColin Angle: iRobot CEO | Lex Fridman Podcast #39
Angle has run iRobot for 29 years and sold over 25 million Roombas, and his honest take is that technology alone never made the company money, he only succeeded once he became, in his words, a vacuum cleaner salesman. He names Anki, Jibo, Mayfield Robotics, and Rethink Robotics as brilliant robotics companies that all went under, and explains why iRobot skipped lidar for cheaper vision-based navigation as cameras got cheap enough. His theory that smarter robots will get more emotional, not more purely logical, is a genuinely different take on where home robotics is headed.
Read the full episode notesSergey Levine: Robotics and Machine Learning | Lex Fridman Podcast #108
Levine's argument reframes the whole field: the gap between humans and robots isn't really a hardware problem, it's an intelligence problem, and robotics is the best tool we have for studying that intelligence rather than something that needs intelligence solved first. He points to a famous 2004 Stanford video of a robot tidying a room and fetching a beer, the punchline being it was entirely teleoperated by a human. His warning that any human-built bottleneck (like a simulator) that doesn't improve from data will eventually become the thing holding the whole system back is worth sitting with if you follow AI safety debates at all.
Read the full episode notesDrago Anguelov (Waymo) - MIT Self-Driving Cars
Anguelov's MIT lecture is the clearest walk-through of why self-driving is hard: not the common cases, but the long tail of rare situations like a cyclist carrying a stop sign or a falling pole. Waymo simulates roughly 25,000 virtual cars driving ten million miles a day, over seven billion simulated miles total, and Anguelov explains how the company deliberately stages dangerous scenarios at a 90-acre former air force base rather than waiting to encounter them on public roads. Good listen for anyone who wants the actual engineering pipeline behind an autonomous vehicle company, not the marketing version.
Read the full episode notesEmilio Frazzoli, CTO, nuTonomy - MIT Self-Driving Cars
Frazzoli, who invented the RRT* motion-planning algorithm and put the first autonomous vehicles on Singapore's public roads, makes a contrarian case that safety isn't even the biggest payoff of self-driving cars, reclaiming people's lost driving time is worth roughly $1.2 trillion a year by his math. He calls the SAE's numbered automation levels an 'enormously bad idea' because levels 2 and 3 require humans to supervise machines, which goes against human nature. His closing thesis, that nobody has ever rigorously defined how a vehicle should behave in the first place, human-driven or not, reframes the entire debate.
Read the full episode notesPieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
Abbeel runs Berkeley's robotics learning lab and built BRETT, the Berkeley Robot for the Elimination of Tedious Tasks, and here he explains why teaching a robot to beat Roger Federer is as much a hardware problem as a software one. His work with Chelsea Finn on letting robots learn from third-person human demonstrations, essentially machine translation for demonstrations, is a genuinely clever idea explained plainly. Good pick for anyone who wants the reinforcement-learning fundamentals without the jargon.
Read the full episode notesEfficient Computing for Deep Learning, Robotics, and AI (Vivienne Sze) | MIT Deep Learning Series
Sze's contrarian insight reframes efficient computing entirely: power in AI systems is dominated by data movement, not the actual computation, and reading from off-chip memory costs two orders of magnitude more energy than the math itself. She walks through how a self-driving car can burn over 2,000 watts just on compute, enough to need water cooling, while a handheld device is capped under a watt. If you want to understand why robots and phones can't just run bigger neural nets forever, this is the clearest explanation of the actual physical constraint.
Read the full episode notesKyle Vogt: Cruise Automation | Lex Fridman Podcast #14
Vogt's path runs from building 200-pound BattleBots as a teenager in Kansas, through the first DARPA Grand Challenge, to co-founding Twitch, and finally landing on self-driving cars as what he calls the greatest applied AI problem of his generation. Cruise had a working highway autopilot prototype within a year and abandoned it entirely to go all-in on fully driverless cars, a bet that shaped the whole industry's direction. Worth hearing for the founder's-eye view of why the retrofit approach to autonomy quietly failed.
Read the full episode notesRodney Brooks: Robotics | Lex Fridman Podcast #217
Brooks co-founded iRobot and Rethink Robotics and has arguably shaped more of real-world robotics than anyone else on this list, and he spends much of this conversation cautioning against extrapolating from flashy AI demos. His account of Rethink Robotics' fatal mistake, getting talked out of a $3,000 plastic-geared robot in favor of $25-35K machines, and the company's later collapse when every buyer with Chinese money failed a national security review, is a rare unfiltered look at how robotics startups actually die. His claim that iRobot's robots helped shut down Fukushima Daiichi because the company had nine years of field experience from 6,500 robots in Iraq and Afghanistan is the kind of detail you don't get from a press release.
Read the full episode notesVijay Kumar: Flying Robots | Lex Fridman Podcast #37
Kumar built a 7,000-pound hexapod in grad school before pioneering the small agile quadrotors that now swarm in formation, and he explains how cheap IMUs, originally developed for car airbags, made that leap possible around 2007-2009. He argues autonomous flight is in some ways easier than driving because a drone can always fall back on a guaranteed-safe trajectory, straight up, across, straight down, that ground vehicles simply don't have. His point that lifting a small UAV costs about 200 watts per kilo, while the entire human brain runs on under 80 watts, sums up why flying cars keep hitting a battery wall.
Read the full episode notesLeslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
Kaelbling started at SRI working on Flakey, the successor to the pioneering robot Shakey, and here she explains 'belief space' planning, the idea that a robot should reason about its own uncertainty over world states rather than just the states themselves. Her account of co-founding the open-access Journal of Machine Learning Research after 75 percent of an existing journal's editorial board resigned over cost and access is a great side story about how academic fields actually change. She refuses to take a dogmatic side in the symbolic-versus-neural-networks fight, arguing abstraction, not ideology, is what makes planning tractable.
Read the full episode notesChris Urmson: Self-Driving Cars at Aurora, Google, CMU, and DARPA | Lex Fridman Podcast #28
Urmson led CMU's DARPA Grand and Urban Challenge teams and later ran Google's self-driving car program before founding Aurora, and he pushes back directly on Elon Musk's claim that lidar is a crutch, arguing lidar, camera, and radar are all essential for real robustness. His warning about level-2 driver assistance is blunt: people sleep in their Teslas, and the economics of driver-assist diverge from the actual path to autonomy. He predicts large-scale driverless deployment within ten years, starting in moderate-speed urban and suburban environments rather than highways.
Read the full episode notesSertac Karaman (MIT) on Motion Planning in a Complex World - MIT Self-Driving Cars
Karaman proved that RRT, the widely used motion-planning algorithm of its era, fails to converge to optimal solutions, and then built RRT* as the fix, a discovery that became his doctoral thesis and now runs quietly inside a huge share of autonomous vehicle software. His retelling of MIT's low-speed collision with Cornell's robot car during the DARPA Urban Challenge, caused by a collision checker that got stuck thinking an obstacle was on top of the car, is a great story about how these competitions actually seeded companies like Google's self-driving program. Good pick for anyone who wants the actual math behind autonomous navigation explained by the person who invented it.
Read the full episode notesSertac Karaman: Robots That Fly and Robots That Drive | Lex Fridman Podcast #97
In this follow-up conversation, Karaman compares Waymo, Tesla, and his own company Optimus Ride, which bets on geofenced areas with humans supervising whole fleets rather than teleoperating individual cars. He frames Waymo as more of a research project building an 'AI engine' than an immediate commercial product, and predicts early deployments will lean on lidar simply because it's easier to build with, even while betting his name on cameras eventually being enough. His closing point about Bellman's equation and the curse of dimensionality, that with 100 variables each taking ten values the possible outcomes exceed the atoms in the universe, is the best one-line explanation of why robotics is hard you'll hear anywhere.
Read the full episode notesThat's fifteen conversations covering home robots, self-driving cars, drones, and the reinforcement-learning theory holding it all together. If any of these got you curious about the guest or the show, browse our full library of episode summaries for the rest of the conversation.