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

Humans and robots cannot plan at the level of raw sensory data and torques for long-horizon tasks; spatial, temporal, and goal abstractions (e.g., ‘afternoon’ vs ‘2:54 pm’, ‘room’ vs precise pose) drastically shrink state spaces and make long-horizon reasoning tractable.

In partially observable settings, agents must control their beliefs—probability distributions over possible world states—so they can decide when to gather information, how to trade off sensing vs acting, and when uncertainty is too high to safely proceed.

Even though optimal solutions for POMDPs are often computationally impossible, formulating problems this way clarifies what’s hard, guides approximation choices, and structures algorithms, rather than pretending the underlying uncertainty does not exist.

There is no single ‘true’ method—symbolic logic, neural networks, model-based RL, and policies are all tools; riding a unicycle, solving algebra, and doing medicine likely require distinct representations, time-space trade-offs, and computational structures.

We lack a clear understanding of the right representational targets for perception in an integrated intelligent agent—beyond steering or labeling images—so progress hinges on discovering structural biases (like convolution) for objects, relations, and higher-level reasoning.