Datasets

 

 

During the experiments, we used two different datasets. These sets differ mainly in the amount of sensor noise.

1. The first dataset was collected during 2.0 hours of robot motion, in which the robot traveled approximately 1,000 meters. This dataset was collected when the museum was closed, and the robot guided only remote Internet-visitors through the museum. The robot's top speed was 50cm/sec. Thus, this dataset was ``ideal'' in that the environment was only sparsely populated, and the robot moved slowly.
2. The second dataset was recorded during a period of 4.8 hours, during which Rhino traveled approximately 1,540 meters. The path of this dataset is shown in Figure 12(a). When collecting this data, the robot operated during peak traffic hours. It was frequently faced with situations such as the one illustrated in Figure 7. The robot's top speed was 80cm/sec.

Both datasets consist of logs of odometry and laser range-finder scans, collected while the robot moved through the museum. Using the time stamps in the logs, all tests have been conducted in real-time simulation on a SUN-Ultra-Sparc 1 (177-MHz). The first dataset contained more than 32,000, and the second dataset more than 73,000 laser scans. To evaluate the different localization methods, we generated two reference paths, by averaging over the estimates of nine independent runs for each filter on the datasets (with small random noise added to the input data). We verified the correctness of both reference paths by visual inspection; hence, they can be taken as ``ground truth.''

 

Figure 13 shows the estimated percentage of corrupted sensor readings over time for both datasets. The dashed line corresponds to the first data set, while the solid line illustrates the corruption of the second (longer) data set. In the second dataset, more than half of all measurements were corrupted for extended durations of time, as estimated by analyzing each laser reading post-facto as to whether it was significantly shorter than the distance to the next obstacle.