IEEE Robotics & Automation Magazine - June 2014 - 51

15

20

10

16

5
y (m)

y (m)

24

12

0

8

-5

4

-10

0

0

4

8

12

16

20

-15
-15

24

x (m)

0
x (m)

(a)

(a)

9

-10

-5

5

10

15

250

8
200
Entropy (bm2 )

Entropy (b)

7
6
5
4
3
2

150
100
50

1
0

0

100

200

300

400

500

600

0
0

100

200

300

400

Time (s)

Time (s)

(b)

(b)

500

600

700

Figure 6. The sample results from the experimental data. (a) A
typical path taken by the robot, starting from the center of the
annulus, is indicated by the solid line, and the final position of
the robot and its sensor footprint are indicated by the dashed
line. The true object location is indicated by the red diamond,
and clutter objects are indicated by green circles. The shaded
cells correspond to the nonzero prior probability of the cell
containing an object. (b) The time history of the entropy of the
distribution of object locations over 12 representative runs.

Figure 7. The sample results from the experimental data. (a) A
typical path taken by the robot, starting from the center of the
annulus, is indicated by the solid line. The true object locations
are indicated by the red diamonds. Shaded cells correspond to
the nonzero prior probability of the cell containing an object.
(b) The time history of the entropy of the distribution of object
locations over ten representative runs.

the prior belief. In this scenario, the robot begins at the center
of the environment and searches for targets in an annular
region surrounding it, where the shaded cells have nonzero
probability of occupancy in the prior. We tested two separate
cases: 1) where there is a single object in the environment and
the robot believes that there are either zero or one target within
the environment and 2) where there are two targets (red dog
bone toys) and the robot believes there are up to four. In the
second case, there were also multiple false targets (dog toys of
varying color and shape) placed within the environment.

In ten runs, the robot correctly located the object within the
precision of the grid. Initially, the entropy decreases slowly as
the robot sweeps out some of the area. The sudden drop is due
to the fact that the number of objects is limited to one; this
causes the distribution to rapidly converge when multiple
detections are made in the same cell as that means that the
object cannot be in any of the other cells. The variation in time
to convergence is due to the random placement of the object,
with short times corresponding to the object being placed
nearer the initial footprint of the robot. The robot failed to
localize the object after a full sweep of the environment in two
runs due to failures in the perception system, which are likely
due to adverse lighting conditions. The system was able to
recover in one such instance, shown by the dark blue line in
Figure 6(b), nearly converging to the incorrect cell before
being switching to the correct cell, causing the large spike in
entropy near the end of the trial.

Single Object
We performed 12 trials with random initial positioning of the
object. Since the total number of RFSs is small (the number of
cells plus the empty set), we elected to use a fixed grid with
1-m resolution. To see the progression of information gain, the
resulting time history of the entropies is shown in Figure 6(b).

June 2014

*

IEEE ROBOTICS & AUTOMATION MAGAZINE

*

51
Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - June 2014
