Jason Fleischer.  Route Communication Between Mobile Robots Using Adaptive Landmark Symbols. Ph.D. thesis,  University of Manchester (U.K.), 2004.

ABSTRACT

This thesis introduces a new multi-robot navigation problem called the route communication task, where two robots give each other directions to target locations in a route environment.  To do this, the robots describe routes to each other using symbols that represent landmarks occurring along the route. The robots learn how to use symbols to represent landmarks without a priori knowledge or human labelling. This thesis aims to investigate the competencies required to perform the route communication task: route description, route following, and symbol learning. These competencies are developed and evaluated individually, and then combined together to study the whole task. The mechanisms for route description are based on landmarks. A method of adaptive, on-line landmark detection is presented that selects as landmarks those locations where an acquired sensory prediction model makes poor predictions, i.e., places where the sequence of perceptions is somehow different than normal. Route following methods are developed that enable a single robot to repeatably follow a route description that it generated itself. Algorithms for learning consistent symbol usage by imitation are developed and evaluated, revealing that learning symbols that represent landmarks has a particular problem not found in simulated symbol learning tasks: the robots are unable to occupy the same location without perceiving each other, and so cannot be sure that they are speaking about the same landmark when trying to imitate each other's symbol usages. This is a particular example of a more general problem in symbol learning that I call the unknown reference problem. Two possible solutions to this version of the unknown reference problem are developed and evaluated. These solutions are in the form of learning biases that allow a robot to probabilistically estimate correct symbol usages for a landmark on a route it is travelling, given another robot's description of the same route. The learning bias algorithms are compared with an attention mechanism that does not suffer from the unknown reference problem. The learning bias method is found to have a lower miscommunication rate than the attention mechanism (which is good), but it also has a low symbol usage mutual information (which is bad). In spite of the success of the methods individually, the results are disappointing when the competencies for route description, route following and symbol learning are combined together to attempt the route communication task. The thesis concludes with a look at what aspects of the system might be responsible for the lack of performance in the route communication task.