This week I was mainly surveying on puplications for recursive EM methods (not restricted to SLAM) in various applications, which is listed below: (1) Online EM algorithm for latent model This paper is more like a tutorial on how to use recursive EM algorithm to estimate parameter and update latent variables.
(2) Recursive EM algorithm for Multi-Target Tracking This paper talks applies recursive EM algorithm to multi-target tracking problem. They formulate E-step by using stochastic approximation inspired by (1) and leave the maximization step unchanged. It can be reduced to standard SLAM in (3) by just a slightly modification.
(3) Online EM algorithm to solve SLAM problem This is the most relevant publication to our work. It uses Sequential Monte Carlo step (like a particle filter) and stochastic approximation (similar in (1)) to construct E-step in their recursive EM algorithm
However, all of these papers I surveyed involved a stochastic approximation step, which requires to construct of "Complete Data Sufficient Statistics", or score vector S, and Fisher Information Matrix, and this is the most difficult part in our implementation I am particularly interested in (or may be confused about) how they construct their score vector and fisher information matrix. None of these papers have proposed a way to construct them.