Interview
Sertac Karaman: Robots That Fly and Robots That Drive | Lex Fridman Podcast #97
- Massive scale autonomous flying is projected to follow resolved autonomous driving deployments, with dense operations in urban airspace expected to first occur with cars due to their contained environment.
- High-density deployment of tens of thousands of autonomous vehicles in major cities like Boston or New York is anticipated to enable shared urban trips costing less than one dollar and personal air transport reducing four-hour journeys to 1.5 hours.
- Future personal transport may utilize "agile airspace" (above stone-throw range but below large aircraft altitude) and involve "perch landing" maneuvers on buildings, while large-scale delivery drone operations are considered a possibility.
- Technology enabling flying saucer-like flight via air ionization and magnetic propulsion has a 50-50 probability of realization within the next 50 years, though traditional airplanes are not expected to disappear quickly.
- Ensuring safety requires software several orders of magnitude more complex than current aircraft systems, relying heavily on machine learning for intelligence and simulation environments that must overcome the difficulty of modeling human behavior.
- A critical challenge involves predicting dynamic human behavior and game-theoretic interactions, a difficulty persisting since 2012-2013 and currently under-addressed by ignoring the ego vehicle's influence on future traffic states.
- Trade-offs between efficiency and livability will become explicit as vehicles are deployed, requiring a balanced approach where aggressive driving improves efficiency but harms livability, while nice driving increases delays.
- Public acceptance is expected to rely on an informed populace capable of making risk choices, with a need to balance innovation against safety through explicit testing and transparent communication.
- Waymo is predicted to treat autonomy as a research project focusing on AI engines and human interaction, while Tesla benefits from market demand for automation features driven by Elon Musk's vision.
- Optimus Ride plans to focus on geofenced environments (approximately two miles by two miles) with 4-6 seat agile vehicles to reclaim parking land, potentially opening up after a stealth period.
- Sensor strategies will likely evolve from initial low-resolution solid-state LiDARs to camera-only systems or LiDARs used solely to reduce compute intensity or aid in safety certification.
- High-throughput computing will enable autonomous drones to exceed human speeds and assist cars with full control or crash avoidance, while machines are expected to consistently beat humans in drone racing due to tireless repeatability.
- Predicting specific product timelines two to three years ahead is deemed unreliable due to technology "time dilation," with iterated learning and experimentation viewed as the primary method for closing technology gaps.
- Despite theoretical computational hurdles such as Bellman's equation and the curse of dimensionality, practical autonomous systems are expected to function effectively in real-world scenarios.