AI research is moving away from ``toy tasks'' such as block-stacking towards more realistic problems. Building autonomous agents that interact with real-world software environments such as operating systems or databases is a pragmatically convenient yet intellectually challenging substrate for AI research. To support this claim, we are utilizing planning and machine-learning technology to develop an Internet softbot ( software ro bot)| a customizable and (moderately) intelligent assistant for Internet users. The softbot accepts goals in a high-level language, generates and executes plans to achieve these goals, and learns from its experience. Custom-built execution and sensing modules enable the softbot to interact with a UNIX shell and the World-Wide Web in real time. We are developing both a graphical user interface and a natural-language interface to the softbot.
Principal Investigators: Oren Etzioni, Dan Weld