Dmitri Dolgov: Waymo and the Future of Self-Driving Cars | Lex Fridman Podcast #147

Dmitri Dolgov traces his path from a kid programming games in 1980s Russia, through Stanford's DARPA Urban Challenge team, to building Waymo's fully driverless service in Phoenix. He explains how the early Google project set audacious milestones (100,000 autonomous miles, 10 hundred-mile routes with zero interventions) and how those lessons fed into a fifth-generation hardware platform with custom lidar, radar, and cameras. He details how Waymo evaluates and deploys its driver, the role of machine learning across the stack, and the three dimensions of scaling beyond Phoenix. The conversation closes on lidar debates, pedestrian safety, the trolley problem, Russian literature, and the meaning of life.

• Waymo's very first day of regular driverless operations in October 2017 included rides with a completely empty driver's seat, not just a safety driver who couldn't intervene.
  00:33:18
• The team initially pursued an L3 freeway driver-assist program before pivoting around 2013 to building a fully driverless vehicle, deciding the two approaches are fundamentally different.
  00:30:43
• Dolgov argues good driving doesn't require aggression; professional limo drivers are smooth, efficient, and assertive without breaking rules, reframing Lex's intuition that you must bend rules to get anywhere.
  01:13:28
• Waymo raised about $3.2 billion in its first external financing round in 2020 despite being part of Alphabet.
  00:38:32
• Responding to Elon Musk's claim that lidar is a crutch, Dolgov rejects it, calling lidar a key complementary sensor and dismissing cost and aesthetics objections.
  01:37:13
• Dolgov rejects the purest end-to-end learning approach (raw sensors straight to steering torque) as 'too much,' favoring a modular hybrid of ML and rule/physics-based components.
  01:42:25
• On the trolley problem, Dolgov says he'd 'kill both,' calling it a red herring; a well-built defensive driver avoids ever ending up in that situation.
  02:02:35

• The DARPA Urban Challenge ran on an abandoned Air Force base built out as a fake city, with autonomous cars sharing roads with professional human drivers and no pedestrians.
  00:15:36
• A bug made Stanford's car take an extra 'victory lap' around an oval at each checkpoint because it wouldn't plan the next waypoint until hitting the previous one.
  00:21:53
• Waymo's fifth-generation vehicles carry 29 cameras, 5 lidars, and 6 radars, mostly custom-designed in-house.
  00:42:45
• Cars operate independently with no required connectivity but share real-time info (accidents, construction) propagated across the fleet as map priors.
  00:48:59
• Phoenix riders pay real money point-to-point; demand exceeds capacity and the youngest riders are one or two years old in car seats.
  00:55:43
• A pickup heuristic once dropped a rider on the wrong side of a parking lot lined with cacti, forcing them to walk around in 110-degree heat.
  01:01:27
• In Phoenix a cyclist on a sidewalk tripped and fell into the car's path; the system's simulated response matched the test driver's correct steer-and-brake reaction.
  01:58:30
• Transformer/language-model breakthroughs (GPT-3 and friends) are being applied to behavior prediction and planning because driving trajectories share sequential, local-plus-global structure with sentences.
  01:53:20
• As a kid in Russia Dolgov wrote a game a Japanese developer offered to buy for a few hundred dollars; he declined and posted it as freeware instead.
  00:03:39
• Lex admits he has read all serious Russian literature in English translation and is reconsidering reading Dostoevsky in Russian.
  02:08:16