The role of Bayesian bounds in comparing SLAM algorithms performance

It is certainly hard to establish performance metrics for intelligent systems. Thankfully, no intelligence is needed to solve SLAM at all. Actually, when we cast the SLAM problem in the Bayesian framework, we already have a formula for the solution — SLAM research is essentially about finding good approximations to this computationally monstrous formula. Still, SLAM algorithms are difficult to analyze formally, partly because of such out-of-model ad hoc approximations. This paper explains the role of Bayesian bounds in the analysis of such algorithms, according to the principle that sometimes it is better to analyze the problem than the solutions. The theme is explored with particular regard to the problem of comparing algorithms using different representations and different prior information.