Fig. 1. The Nomad 200 mobile robot. A flux gate compass was used to keep the turret, and therefore the sensors, at a constant orientation.

Fig. 2. (a) Raw odometry. (b) Compass-based odometry. The accumulated rotational drift in the robot’s raw odometry was removed on-line using the compass sense.

Fig. 3. Example occupancy grid and histograms. Occupied cells are shown in black, empty cells in white and unknown cells in grey.

Fig. 4. Matching the x and y histograms. The new histograms are convolved with the stored histograms for each place in the robot’s map to find the best match.

Fig. 5. Left: environment C in Table 1. Right: retrospectively corrected odometer data used for performance evaluation (see also Fig. 2).

Fig. 6. Global localisation experiment. This shows the global localisation performance of the new self-localisation system when starting from an unknown location, and then being kidnapped after 30 m of travel. The performance is compared to that of occupancy grid matching, as in Table 2, and global localisation using the baseline method for scan matching described in Section 5.2.

Table 1. Characterisation of environments (the number of places N in the respective maps and the number of trials used for testing in each environment are also shown)

Table 2. Scan matching experiment (this shows the percentage of correct localisation attempts for each environment and the mean value μ over all four environments; the mean processor time, T, to match one pair of scans is given in units of t=1.83×10⁻⁵ s, as measured on a Sparcstation 20)