Interview
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
- Leslie Kaelbling, a roboticist and MIT professor, cites Gödel, Escher, Bach as the catalyst for her interest in AI, specifically the concept of generating complex behavior from simple primitives.
- Her transition from philosophy to robotics occurred after being hired by SRI to work on a robot following the Shakey project, where she had no prior background in control or sensors.
- Kaelbling views her philosophy background as highly relevant to AI, particularly regarding formal notions of belief, knowledge, and denotation, which she considers one step removed from computer science logic.
- She identifies as a materialist who believes robots can be behaviorally indistinguishable from humans, though she remains agnostic regarding the philosophical "zombie" question of internal consciousness.
- She attributes the failure of 1980s expert systems to the flawed assumption that humans can articulate their decision-making processes into logical statements.
- Kaelbling argues that the "symbolic vs. neural network" debate is a false dichotomy, asserting that different reasoning styles (symbolic abstractions, neural networks) are necessary for different problem domains.
- She advocates for the use of abstractions (spatial, temporal, and goal-based) to reduce the state space and horizon required for complex planning, such as navigating an airport or earning a PhD.
- Her work emphasizes hierarchical planning, where high-level abstract plans are made with a "leap of faith" regarding feasibility, relying on learned models to predict the difficulty of unexecuted sub-tasks.
- Markov Decision Processes (MDPs) model uncertainty in future transitions but assume full state observability, which is rare in real-world robotics.
- Partially Observable MDPs (POMDPs) are used to model situations where the agent does not know the true state and must reason based on a history of noisy observations.
- Kaelbling notes that optimal planning for even discrete POMDPs can be undecidable, necessitating approximations rather than optimal solutions.
- She advocates for the concept of "belief space," where an agent plans by controlling its own uncertainty and choosing actions specifically to gather information (active sensing).
- Current AI research faces a methodological crisis where engineering leaps outpace scientific theory, leaving researchers unable to formally prove why deep learning systems work.
- Kaelbling suggests the next breakthrough in perception requires building in structural biases (similar to convolutional layers) to reduce sample complexity and define the desired output of perception systems.
- She dismisses the need for consciousness or self-awareness as prerequisites for advanced robotics, suggesting a system with internal monitoring of its components suffices for functional "zombies."
- She views the current publishing model in AI as unsustainable due to the volume of papers, preferring curation and public commentary over traditional peer review or "upvote" systems.
- Kaelbling expresses concern that the academic incentive structure favors short-term publication outputs, disincentivizing the long-term, deep work required to solve hard problems.
- She predicts AI development will follow a cyclical pattern of hype and winter, but with each cycle reaching a higher technological "high watermark."
- Regarding safety, Kaelbling argues that the primary risk is not robot autonomy but "objective misalignment," where the optimization of a poorly defined goal function leads to unintended outcomes.
- She distinguishes between programming (step-by-step commands) and modern AI (optimizing objective functions over a hypothesis class) to correct public misconceptions about robot unpredictability.
- Her most compelling research question is determining the optimal balance between built-in structural knowledge and data-driven learning to create robust, real-world robotic systems.
- Kaelbling does not have a favorite science fiction robot, stating she cares more about the engineering process itself than the final production of a specific character.