Sertac Karaman: Robots That Fly and Robots That Drive | Lex Fridman Podcast #97

Sertac Karaman discusses why deploying autonomous vehicles at scale in human-present environments is far harder than the underlying technology suggests. He compares the strategies of Waymo, Tesla, and his own company Optimus Ride, which focuses on geofenced areas with humans supervising fleets rather than full teleoperation. The conversation covers simulation, the difficulty of modeling human behavior, the lidar-versus-cameras debate, and high-throughput computing for aggressive drone flight. Karaman closes on Bellman's equation and the curse of dimensionality as the most beautiful idea in robotics.

• Karaman bets there is a 50/50 chance flying-saucer-like machines that ionize air and push it down with magnets could exist within 50 years.
  00:07:20
• Optimus Ride's core vision is moving from one person per vehicle to ten people operating fifty vehicles via high-level supervision, not teleoperation.
  00:31:46
• He frames Waymo as more of a research project building an 'AI engine' rather than an immediate product.
  00:47:53
• Despite Optimus Ride using lidar, Karaman says he'll put his name on a future where cameras alone are enough for autonomous driving.
  00:59:23
• He predicts early self-driving deployments will rely on lidar because it is simply easier to build with.
  01:02:30
• On drone racing: run a course once and machines lose to humans, but run it ten times and machines win consistently as humans tire.
  01:16:35

• It costs 2.4 cents to produce a single penny, with roughly 85 million dollars spent annually making them.
  00:01:01
• Humans easily spot rendered fakes only when a human figure is in the scene, suggesting our brains are specially trained to recognize people.
  00:13:33
• In a two-mile-by-two-mile area, reclaiming parking land could mean tens of millions in savings, or billions in central New York.
  00:38:31
• Shuttles are hated because drivers are expensive, pushing operators toward big 20-30 seat vehicles that are slow and inflexible.
  00:40:05
• Karaman's group builds vision systems that see at a kilohertz, while the human eye barely reaches a hundred hertz.
  01:05:38
• Camera frame rates are capped by the Shannon limit on copper wires, a fundamental data-transfer bottleneck.
  01:08:43
• Chips can't be clocked much faster partly because at 3 GHz light only travels about an inch per clock cycle.
  01:09:44
• With 100 variables each taking ten values, the possible assignments exceed the number of atoms in the universe.
  01:20:14