Robotics, AI Planning, and Intelligence with Leslie Kaelbling

The Intersection of Philosophy and Robotics

Leslie Kaelbling explores how her background in philosophy at Stanford provided a crucial framework for her work in artificial intelligence. While she acknowledges the philosophical debates regarding consciousness and the "zombie" problem, she maintains a materialist perspective, viewing the technical challenges of creating human-like machines as primary.

The Evolution of Robotics and AI

Kaelbling discusses the historical oscillation of AI trends, from early cybernetics to expert systems and finally modern reinforcement learning.

• Shaky and Flaky: Reflecting on her early days at SRI, she highlights the importance of learning by "reinventing the wheel," which allowed her to better appreciate the complexities of robot control and sensors.
• The Limitations of Expert Systems: She explains that the downfall of early expert systems was the false assumption that humans could introspectively articulate their own reasoning processes.

Rethinking Planning and Abstraction

A major portion of the discussion focuses on how agents navigate uncertainty.

• Belief Space vs. State Space: Kaelbling argues that instead of just controlling physical states, a robot must learn to manage its own beliefs (probability distributions) about the world to perform intelligent information gathering.
• Hierarchical Planning: To manage long-horizon tasks—like traveling through an airport or earning a PhD—she advocates for hierarchical reasoning. By using abstractions, agents can make necessary leaps of faith without needing low-level details for every sub-step.

The Future of AI Research

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

Kaelbling emphasizes that the field needs a shift from mere empirical "hacking" toward more rigorous science. She advocates for finding the right balance between learning (neural networks) and engineered structure (built-in biases), similar to how convolution serves as a structural bias in spatial perception.