I work on building autonomous agents that could understand high-level language commands. My goal is to create AI agents that would free human beings from tedious tasks and aid them in better decision makings.

We proposed an intuitive formalism for representing procedures as programs and subsequently applied this concept to broader tasks with large language models (PaL, CoCoGen). We built the first large-scale hierarchical procedural knowledge base. To learn from the knowledge base and generate new and previously unseen procedures, we designed DocPrompting that reads the relevant documentation before taking actions. With the belief of “what I don’t measure, I can’t improve”, we built WebArena, a realistic and reproducible environment for building and evaluating autonomous agents that are guided by high-level natural language commands.