AML Talk

Despite many advances in human-computer interaction, most development systems require that users transfer domain knowledge to computers and robots through traditional procedural-programming techniques. However, most users have neither the expertise nor the inclination to create programs to perform their tasks. In industrial settings, many companies do not have the resources to automate production. For example, ABB Flexible Automation estimates that 98\% of arc welding is performed by hand due, in part, to the complex and time-consuming programming process. The down-time required to re-program the facilities may interrupt production. Furthermore, the expense required to obtain programming expertise (in terms of consultants or training) may be too great.

The aim of this research is to create a system that increases the knowledge transferred between users and computers or robots. This system should simplify the programming process and increase user productivity. To ensure that domain knowledge is conveyed in the most effective manner, observations obtained from user demonstrations serve as input to generate and synthesize the programs. The primary issues to address in this research involve how to create programs, which are inherently deterministic, from the actions of users, which can be ambiguous, non-repeatable, and unintentional. For any system to succeed under the noisy nature of these conditions, it must be able to cope with the inevitable uncertainty that arises. Thus, programs that the system generates must not merely mimic the actions of the user, but the programs should capture the \emph{intent} of the user.

This paradigm of generating programs from user demonstration is broadly called \emph{Learning by Observation (\emph{LBO}), Programming by Demonstration, Teaching by Example}, or some permutation thereof.

In this talk, I will describe the approach we have taken and where we are going. This is going to be a miniature version on my PhD proposal talk, so I'm really looking for feedback.