Monte Carlo Localization for Teach-and-Repeat Feature-Based Navigation

This work presents a combination of a teach-and-replay visual navigation and Monte Carlo localization methods. It improves a reliable teach-and-replay navigation method by replacing its dependency on precise dead-reckoning by introducing Monte Carlo localization to determine robot position along the learned path. In consequence, the navigation method becomes robust to dead-reckoning errors, can be started from at any point in the map and can deal with the ‘kidnapped robot’ problem. Furthermore, the robot is localized with MCL only along the taught path, i.e. in one dimension, which does not require a high number of particles and significantly reduces the computational cost. Thus, the combination of MCL and teach-and-replay navigation mitigates the disadvantages of both methods. The method was tested using a P3-AT ground robot and a Parrot AR.Drone aerial robot over a long indoor corridor. Experiments show the validity of the approach and establish a solid base for continuing this work.