Skip to content
UoL CS Notes

Grid Localisation

COMP329 Lectures

This is where we model the world as a 2D or 3D array. We can then store the probability that we are in any one location in this array.

Piecewise Constant Representation

Markov localisation utilised a probability distribution across all states:

  • Represented as $\text{Bel}(x_t)$ for each $x_t$.

We can model this discretely as a histogram (bins):

  • If there are $N$ rooms, then we can model this as:

    \[p(n)\forall n\in N\]

The sum over all bins should be:

\[\sum x_t=1\]

We can then calculate like so:

  1. Start with a uniform distribution (if the location is unknown), or bounded Gaussian (if a location is known).
  2. The robot observes a feature:
    • The Initial belief is proportional to the observation likelihood.
    • The histogram then represents likely locations of the robot using:

      \[P(z_t\mid x_t)\]
  3. The robot now moves:
    • We convolve out current distribution with our motion model.
    • This translates the histogram and smooths peaks.
    • This is represented as:

      \[\sum_{x_{t-1}}P(x_t\mid u_t,x_{t-1})Bel(x_{t-1})\]
  4. The final distribution is the product of the observation model and resulting distribution with the shifted motion model:

    \[Bel(x_t)=\eta P(z_t\mid x_t)\sum_{x_{t-1}}P(x_t\mid u_t,x_{t-1})Bel(x_{t-1})\]

    $\eta$ is a normalisation factor that ensures $\sum x_t=1$.