Naman Shah

I am a Research Scientist at the Allen Institute for AI (Ai2), working with Dieter Fox. Before Ai2 I was a postdoctoral researcher at Brown University with George Konidaris, and I completed my Ph.D. at Arizona State University with Siddharth Srivastava.

My work asks what a robot should represent about the world in order to act in it, and how it can arrive at that representation on its own. The decisions that matter, such as which object to pick up or when a grasp has failed, live at a coarser level than sensors and controls, and that level is usually hand-written. I learn it instead: state and action abstractions built from a robot’s own experience, with soundness guarantees that carry over to the planners built on them. The same machinery makes reinforcement learning composable, turning a single learned skill into hours-long behavior on real mobile manipulators and bimanual arms.

At Ai2 I am scaling this up by generating robot interaction data across thousands of GPU-accelerated simulation environments, and training hierarchical vision-language-action models that decompose a task into subgoals and realize them through learned control.