Leslie Kaelbling argues robotics is unsolved and compositionality may help

Leslie Kaelbling says robotics still has hard open problems, and that compositionality and modularity may help solve them.

In an IROS 2026 workshop talk titled "Small models for big worlds," she says Astra performs well at robotics partly because of a compositional natural-language model of the world and recursive tool calls. The notes from the talk also list what an intelligent autonomous robot still needs, including resource-rational computation and persistent memory.