Simulating in a physical world

The two different maps used in the simulation for the MCL2D algorithm are also used in the simulation in a physical world. The maps were printed in the size 841 mm x 594 mm on a plotter.
The algorithm is calibrated on the used gray scales. The robot is placed on the map with a relative direction of zero degrees and on several, different positions. In these first experiments the robot is not interacting with the world. The world is also static. To follow the test we put a picture of the physical world and the map representation with samples together. As it is obviously, by knowing about the relative direction, the first map is still unique, the second is not.

The first map:
The algorithm is performing an amount of rounds before it is obviously that the samples will build a cluster at a certain position on the map. After all the tests there was always a cluster at the right position which wins after several steps.
The test show that the MCL2D algorithm works quiet accurate in the physical environment. Most of the approximations don't have an strong effect on the error.


In this experiment the robot car is set in the right lower corner. The MCL2D algorithm developed to cluster of samples. The cluster at the right position is growing while the other one is shrinking. After a certain amount of steps the algorithm was able to localise the robots position.


In this experiment the robot car is set in the right upper corner. The MCL2D algorithm developed to cluster of samples at the two upper corner of the map. The cluster at the right position is growing while the other one is shrinking. After a certain amount of steps the algorithm was able to localise the robots position.


In this experiment the robot car is set in the left upper corner. In this case the MCL2D algorithm doesn't developed different clusters of samples. The cluster at the right position is growing quickly. After a certain amount of steps the algorithm was able to localise the robots position.

The uniqueness of the map helps the algorithm to find the right point without following a trajectory.

The second map:
The ambiguity of the second map is obviously a problem for the algorithm. To solve this problem the robot has to change its position at one point in the experiment. After all it's possible to say that the MCL2D algorithm works well on the second map. There was always a cluster at the right position with a strong probability and a lot of samples. Most of the chosen position where localised in the experiments


In this experiment the robot car is set in the right upper corner. The MCL2D algorithm developed two clusters of samples at different positions with approximately the same expected sensor readings. The cluster at the right position is growing while the other one is shrinking, but it took a lot of steps in comparison to the first map. After a certain amount of steps the algorithm was able to localise the robots position.


In this experiment the robot car is set in the left lower corner. The MCL2D algorithm developed two clusters of samples at different positions with approximately the same expected sensor readings. The clusters have a big concurrency so no one of them is able to get all the samples.
It's necessary to drive a trajectory with the robot to get more information about the actual position and to localise the robot.

A trajectory is needed to solve the ambiguity of the map.