Markov Localisation

COMP329 Lectures 2022-10-27

This is the process of using landmarks to reduce the uncertainty of a robot’s location.

1. We start with some distribution across all states:

   • This includes heading as well as position.

   • This is typically uniform if the initial robot position is unknown.

     This means that we don’t know where we are.

2. This distribution tells us the probability of the robot begin at each state:

   The belief is proportional to the observational likelihood.

3. As the robot observes the environment, it spots a landmark:

   • The distribution is updated to reflect a Guassian aroudn all instances of the landmark.
   • Other probabilities will fall, so that total across the state space is equal to one.

4. The robot then moves:

   • Find the convolution of the probability distribution with the motion model.

     This shifts the distribution model in the direction of the move and smooths the distribution.

5. A new observation is made:

   • We multiply the distribution from the movement with that of the observation to update the belief.