Personal Robot Training via Natural-Language Instructions.

Future domestic robots will need to adapt to the special needs of their users and to their environment. Programming by natural language will be a key method enabling computer language-naive users to instruct their robots. Its main advantages over other learning methods are speed of acquisition and ability to build high level symbolic rules into the robot. This paper describes the design of a practical system that uses unconstrained speech to teach a vision-based robot how to navigate in a miniature town. The robot knows a set of primitive navigation procedures that the user can refer to when giving route instructions. A particularity of this project is that the primitive procedures are determined by analysing a corpus of route instructions. It is found that primitives natural to the user, such as “turn left after the church” are very complex procedures for the robot, involving visual scene analysis and local route planning. Thus, to enable natural user-robot interaction, a high-level of intelligence needs to be built into “primitive” robot procedures. Another finding is that the set of primitive procedures is likely not to be closed. Thus, on time to time, a user is likely to refer to a procedure that is not preprogrammed in the robot. How best to handle this is currently investigated. In general, the use of Instruction-Based Learning (IBL) imposes a number of constraints on the design of robotics systems and knowledge representation. These issues and proposed solutions are described in the paper.