Catalogue of Artificial Intelligence Techniques
Aliases: Position Estimation
Keywords: bayes, localization, markov
Author(s): John Hussey
Markov localization is a technique that allows the estimation of a robotís position from sensor data and a map. This works by maintaining a probability distribution over all possible robot poses. The probability represents the robotís belief that it is in a certain location, where belief is a degree of confidence about the robots locality. Markov localization addresses the problem of global localization and position tracking but only in a static environment. It encounters problems when working in a dynamic environment.
Markov localization works by taking the robotís belief at time t-1, the robotís sensor data, the robotís control data and a map, then it returns the robotís belief at time t. A probability distribution is maintained over all possible robot positions, the distribution representing the robots belief about where it is on the map. Modes in this distribution represent possible locations of the robot. A unimodal distribution would suggest that the robot has a high degree of confidence about its location, whereas a multimodal distribution suggests that the robot is quite uncertain about its location. In the case of the global localization problem, where the robot has no idea where it might be, the probability distribution is uniform.
However Markov localization is not robust in dynamic environments. For example if the robot is moving through a room of crowded people, it is going to have trouble matching its sensor readings to a position on its internal map, as the people will be obfuscating sensor readings and disrupting an otherwise reliable probability distribution. More practical techniques are required to deal with localization in a dynamic environment in real time.
Markov localization may be successful in static environments, but it does rely on the assumption that the robotís location is the only state in the environment that systematically affects sensor readings. This is why it comes unstuck in dynamic environments.
- Sebastian Thrun, Wolfram Burgard, Dieter Fox, Probabilistic Robotics (Ronald C. Arkined.), MIT Press, 2005.