Making Robot Learning Controllable: A Case Study in Robot Navigation

In many applications the performance of learned robot controllers drags behind those of the respective hand-coded ones. In our view, this situation is caused not mainly by deficiencies of the learning algorithms but rather by an insufficient embedding of learning in robot control programs. This paper presents a case study in which ROLL, a robot control language that allows for explicit representations of learning problems, is applied to learning robot navigation tasks. The case study shows that ROLL’s constructs for specifying learning problems (1) make aspects of autonomous robot learning explicit and controllable; (2) have an enormous impact on the performance of the learned controllers and therefore encourage the engineering of high performance learners; (3) make the learning processes repeatable and allow for writing bootstrapping robot controllers. Taken together the approach constitutes an important step towards engineering controllers of autonomous learning robots.