Leslie Kaelbling on Reinforcement Learning, Planning, Robotics, and the Philosophy of AI

Leslie Kaelbling, a leading roboticist, discusses her journey into AI and robotics, emphasizing the foundational role of philosophy and logic in her work. She argues that the current limitations of AI and robotics are primarily technical, not philosophical, maintaining a materialist view that intelligent machines are fundamentally achievable. A core argument is the necessity of abstraction and hierarchical planning for tackling complex, long-horizon problems in AI, moving beyond fine-grained control to more conceptual reasoning. Kaelbling distinguishes between the use of logic by engineers for system analysis versus its direct manipulation within a robot's "head," as seen in the situated computation approach. She highlights the historical oscillation in AI research, where problems are often abandoned or reframed when current methods fail, rather than being solved. A crucial nuance is her stance on expert systems, acknowledging the utility of symbolic reasoning and abstraction while criticizing the flawed assumption that humans can effectively articulate their own decision-making processes for encoding. She also differentiates between Markov Decision Processes (MDPs), which assume full state observability, and Partially Observable Markov Decision Processes (POMDPs), which account for real-world uncertainty and incomplete information. For designing intelligent systems, Kaelbling advocates for a multi-faceted approach, rejecting dogma in favor of applying diverse reasoning styles (continuous, discrete, symbolic, neural) as dictated by the problem. She stresses the importance of "reinventing wheels" for deeper understanding before adopting established solutions. Her work on belief space planning offers a practical framework for agents to deliberately gather information and manage their own uncertainty, optimizing not just world state but also their understanding of it. For intractable problems like optimal POMDP planning, she recommends embracing approximation in both modeling and algorithms, focusing on "doing a very bad job of very big problems" to make progress. The conversation touches upon the broader implications for the future of AI, particularly the current "methodological crisis" where engineering advancements outpace scientific understanding. Kaelbling calls for a return to theoretical rigor and formal solution concepts to transform empirical successes into predictable science, akin to how bridge building evolved from trial-and-error to predictive engineering. Her insights into hierarchical planning and abstraction suggest a path towards more human-like intelligence in robots, capable of navigating complex, uncertain environments by making "leaps of faith" and generalizing from past experiences, ultimately pushing the boundaries of what autonomous systems can achieve.

I read Gödel, Escher, Bach when I was in high school that was pretty formative for me because it exposed uh the interestingness of Primitives and combination and how you can make complex things out of simple parts and ideas of AI and what kinds of programs might generate intelligent Behavior.

the parts of philosophy that are closest to AI I think or at least the closest to AI that I think about are stuff like belief and knowledge and denotation and that kind of stuff and that's you know it's quite formal and it's like just one step away from the kinds of computer science work that we do kind of routinely.

at least my personal view is that I'm completely a materialist and I don't think that there's any reason why we can't make a robot be behaviorally indistinguishable from a human.

everybody should read the shaky tech report because it has so many good ideas in it I mean they invented a star search and symbolic planning and learning macro operators they had uh low-level kind of configuration space planning for their robot they had Vision they had all this the basic ideas of a ton of things.

I'm big in favor of wheel reinvention actually I mean I think you learned a lot by doing it yes uh it's important though to eventually have the pointers to so that you can see what's really going on but I think you can appreciate much better the the good Solutions once you've messed around a little bit on your own and found a bad one.

the main road block I think was that the idea that humans could articulate their knowledge effectively into into you know some kind of logical statements so it's not just the cost the effort but just the capability of doing it.

my biggest point about all of this is that there should be that Dogma is not the thing right we shouldn't it shouldn't be that I in favor against symbolic reasoning and you're in favor against neural networks it should be that just just computer science tells us what the right answer to all these questions is if we were smart enough to figure it out.

I like to you know I like to say that I I'm interested in doing a very bad job of very big problems.

my control problem if I'm reasoning about how to move through a world I'm uncertain about my control problem is actually the problem of controlling my beliefs so I think about taking actions not just what effect they'll have on the world outside but what effect I'll have on my own understanding of the world outside.

you have to have you have to make a leap of faith that you can figure it out once you get there and it's really interesting to me how you arrive at that how do you so you have learned over your lifetime to be able to make some kinds of predictions about how hard it is to achieve some kinds of sub goals.