Learning to Traverse Doors Using Bayesian Networks

By: Elena Lazkano and Basilio Sierra

Mobile robots need to navigate in their environment in order to perform useful tasks. Doors appear in almost every office-like indoor environment and often doors have to be crossed during the navigation process. We believe that visual information may help to anticipate that a door has to be crossed and that the visual information could be combined with proximity sensors in order to select a good position from which the door crossing behavior could start.

This work presents the experiment we are carrying out with a B21 mobile robot. The main objective of the experiment is to use the images taken from the CCD camera mounted on the top of the robot in order to detect and cross opening doors. We divide the traversing doors behavior in two steps:

1. door identification and aproximation
2. door crossing.

For the first step, we consider that doors are parallel vertical lines separated by a minimum distance. Thereby, we use a Sobel filter to detect vertical lines in the image. Once the robot detects a posible matching, she makes the necessary movements to get closer to the door until a preventive distance, where the door is still visible and there is no collision danger, is reached. This aproximation also helps the robot to reject some false doors. From this position the robot learns the actions it has to perform in order to cross the door (2nd step).

To learn the actions to do, we use as input data the sensor readings and the position of the door lines in the image. We compare the learning results obtained with a Bayesian Network (BN) with several machine learning paradigms in order to select the method that best adapts to the problem. In this experiment, the magnitudes of rotational and traslational velocity during the door crossing step is fixed. Only the signs of those variables vary from action to action. The results obtained by the Bayesian Network in the first experimental phase are the bests among all paradigms used, so we have decided to use BNs in the robot. In order to do the inference, we use the HUGIN software. By this way, we obtain in each moment the action the robot should do given the values of the sonars it has.

This is a real application that shows the use of Bayesian Networks as Supervised Classifier paradigm. The BN learning have been done using three score+search methods, using the metrics K2, BIC and Entropy. We have also used a fixed BN structure in order to compare the automatic learning proccess with the use of an expert knowledge to construct the BN.