How does it converge ?

Particles are created, randomly spreaded all over the room with an orientation centered on X axis with a gaussian distribution (sigma adjusted with envirnoment bias characteristics).

The loop

The robot is oriented following the X axis ( same orientation as during the learning step).

Figure 5 created oriented particles (dot)

The deep learning system returns the most likely positions according to the scan. Particules probabilities are updated following a gaussian distribution centered on each of the most likeky positions and particules filter is resampled.

Figure 6 Particles have been resampled P1 the most likely position

Robot is moved (rotation, distance) . Particles are moved with a function that is adjusted with physical bias (gaussian distribution on rotation and distance). P1 will be moved to P’1.

Figure 7 Particle have been resampled P'1 the new P1 position

The deep learning system returns the most likely positions according to the scan. Particules probabilities are updated following a gaussian distribution centered on each of the most likeky positions P2 and particules filter is resampled. After a few iterations if P2 remains close to P’1 it is highlikely that P2 is close to the actual robot position.

Figure 8 Particles have been resampled P2 the new estimated position