Scenario 3. Several landmarks

    

Description: Scenario with many landmarks and with no obstacles apart from the landmarks themselves. In order to have an interesting scenario, we placed the target landmark label higher, so that it was visible from the starting point, even if there were other landmarks in the view line from the robot to the target. If we had not done so, the path from the starting point to the target would always have been clear, since the target has to be initially visible, which actually corresponds to the first scenario. This change required the robot to move the camera up and down to be able to have the target landmark in its view field (in the previous scenarios, it was only doing a pan movement, with no tilt at all). Thus, we had to change the looking actions in order to incorporate the tilt angle. The agents bidding for looking actions added the tilt angle in the following way: the Target Tracker selects a random tilt angle, ranging from 0 degrees (so that the target landmark can be in the view field when it is 7-8 meters away) to 35 degrees (so that the target can be in the view field when it is less than 1 meter away); the Risk Manager does a similar thing, but it only selects a random tilt angle on one third of the actions it bids for, while it sets a null tilt angle on the other two thirds, since most of the landmarks (actually, all but the target) are at the same height of the cameras (i.e. in the null tilt angle plane); finally, the Rescuer, when bidding because the imprecision is too high, does two visual scans around the robot, one with a null tilt angle, and another one with a random positive tilt angle.

Task: Reach the target landmark, eventually avoiding others along the way and build a map of the environment.

Results: The behavior of the robot in this scenario was similar to the one exhibited in the previous one. However, it reached the target in 85% of the trials; in 15% of the trials it failed because the error on the location of the target made it suppose it was at the target location when it was really not there yet. This was caused by the target being occluded by other landmarks, and the constant change in trajectory needed to avoid these landmarks. These two factors caused the location of the target stored in the Visual Memory to increase its imprecision. However, the imprecision was not high enough for the Rescuer to become active. A difference with the previous scenario is that the Risk Manager bid for looking both ahead and around, since there were many landmarks, and at some point, it had enough landmarks ahead, but not around, so it bid to look around. Some examples of maps built in scenarios of this class during the trials are shown in Figure 6.8. In these maps, numbers represent landmarks the robot has seen, and the triangular regions correspond to topological units of the Map Manager's topological map (see Chapter 3 for details on how this map is built).

Figure 6.8: Maps of 2 different scenarios of scenario class 3