Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15

Leslie Kaelbling traces her path from a Stanford philosophy degree into AI and robotics, starting at SRI working on the robot Flakey (successor to Shakey). She argues against dogmatic positions on symbolic reasoning versus neural networks, emphasizing that abstraction (spatial and temporal) is what makes long-horizon planning tractable. She explains Markov decision processes, partially observable MDPs, and the powerful idea of reasoning in 'belief space' to plan deliberate information-gathering. She discusses the field's methodological crisis where engineering races ahead of theory, the founding of the open-access JMLR, and her worry that short publishing horizons discourage deep multi-year research. On AI's future she dismisses fear of a robot apocalypse but stresses the importance of objective functions and value alignment.

• Kaelbling fell in love with AI by reading Gödel, Escher, Bach in high school, then got into robotics via her first job at SRI's AI lab.
  00:01:05
• She had zero background in robotics, control, or sensors and 'reinvented a lot of wheels' getting the Flakey robot to work.
  00:08:28
• She introduces 'belief space' planning: a robot controls its beliefs (probability distributions over world states), not just the world state.
  00:26:41
• She embraces intractability, saying optimal planning for even discrete POMDPs can be undecidable but that's exactly what makes AI hard and interesting.
  00:22:00
• She co-founded JMLR after roughly 75% of the Machine Learning journal editorial board resigned over cost and access complaints.
  00:45:18
• In her thesis-era day, students did NOT publish papers during their PhD; they chewed on one hard problem for years before writing.
  00:53:13
• Military ethicists revealed they thought robots are programmed like LEGO Mindstorms, prompting her new mission to teach how ML abstraction really works.
  00:55:22

• Kaelbling did an undergraduate degree in philosophy at Stanford because there was no computer science undergraduate degree there at the time.
  00:01:36
• The Shakey robot was sitting in a corner dripping hydraulic fluid into a pan by the time she arrived at SRI.
  00:05:50
• Shakey navigated by vision against black-painted baseboards, localized itself in a map, detected surprising objects, and replanned.
  00:06:20
• She reframes reinforcement learning rewards as 'pleasures,' calling it a more fun name for the concept.
  00:08:58
• She dislikes the term 'symbolic' because she doesn't know what it means technically, but firmly believes in abstraction.
  00:15:12
• She uses walking through an airport as her favorite example: you plan to fly somewhere without knowing your gate or the people in front of you.
  00:30:48
• For a long time JMLR had no bank account and cost only a couple hundred dollars a year, mainly a lawyer and IP address.
  00:46:55
• She floated an idea for 'Leslie's friends who review papers,' a public review group that finds and rates papers without publishing them.
  00:50:35
• Kaelbling has no favorite science-fiction robot; she says she does research because the engineering process is fun, not for what it produces.
  01:00:34