Evidential FastSLAM for grid mapping

We present a solution to the problem of simultaneous localization and mapping (SLAM) based on Dempster-Shafer theory. While several works on the mapping problem based on belief functions exist, none of these approaches deal with the full SLAM problem. In this paper, we derive an evidential version of the FastSLAM algorithm based on a Rao-blackwellized particle filter where belief functions are used for representing a grid map of the robot's environment. The resulting algorithm includes the probabilistic FastSLAM solution as a special case without changing its computational complexity. Due to the additional dimensions of uncertainty provided by belief functions, generated maps explicitly show missing information and conflicting sensor measurements.We evaluate our approach using a simulated robot with sonar sensors, for which we derive evidential forward and inverse models. We compare maps obtained by different combination rules and show that the evidential solution outperforms the Bayesian one regarding the resulting localization error.