FastSLAM

FastSLAM is a class of simultaneous localization and mapping (SLAM) algorithms that use a Rao–Blackwellised particle filter to jointly estimate a robot’s path and a map of its environment. In FastSLAM, the posterior p(X, m | z, u) is factorized as p(X | z, u) ∏ p(m_n | X, z), meaning that given a trajectory hypothesis X, landmark estimates are conditionally independent. Each particle therefore represents a trajectory hypothesis and carries its own map, consisting of mean and covariance estimates for each landmark.

The algorithm operates online as the robot moves. A set of particles is propagated through a motion model using the control u; weights are updated according to the likelihood of observed measurements z given each particle’s pose. For each visible landmark, an incremental update is performed with an individual Kalman filter inside the particle’s map. Data association is performed per particle, selecting which landmark the observation corresponds to or creating a new landmark. After processing measurements, particles are resampled according to their weights, reconstituting a distribution over poses and maps.

History and variants. FastSLAM was introduced by Montemerlo, Thrun, Koller and Wegbreit in the early 2000s as an efficient alternative to full SLAM with a large factorization exploited by the Rao–Blackwellization. FastSLAM 2.0, published later, improves data association and resampling to achieve better accuracy and consistency, often with fewer particles.

Advantages and limitations. FastSLAM reduces the computational burden by exploiting landmark independence conditioned on the trajectory, yielding complexity linear in the number of observed landmarks per particle. It scales well with map size compared to prior methods but requires a sufficient number of particles to cover the trajectory space; incorrect data association or highly dynamic environments can degrade performance. Memory usage grows with the number of landmarks stored in each particle's map.