We discuss different approaches of self-localisation in the Simulation League. We found that the properties of the soccer server’s quantization function tend to produce artifacts with the common approaches, which we try to deal with using a new method: We simulate the player’s position, and dynamically correct this estimate with a gradient descent function by the minimal amount necessary to make it consistent with the perceived flag positions (where we allow for error margins according to the quantization function). It can be shown that self-localisation using the new method is not only relatively exact (between 6.6% and 35% smaller average error than the ’Nearest Flag’ algorithm) but also yields much more consistency than the other approaches....