Information Technology Reference
In-Depth Information
to execute the whole plan prefix prior to the execution of knowledge acquisition tasks.
However, due to the fact this is not always the best strategy it is possible to specify
domain specific control rules.
4
Experimental Results
The proposed plan-based control system is implemented and evaluated on the mobile
service robot TASER (see Fig. 5(a)) as well as using a set of simulated domains.
4.1
Physical Service Robot
This chapter describes an experiment that was conducted with the service robot TASER.
The robot had the task to clean a table. Cleaning a table includes the execution of several
typical service robotic tasks including: pick up an object from a table, navigate to a
desired goal position, find a garbage can, throw away objects, and pick up a garbage
can.
Initially, the robot had no knowledge about dynamic aspects of the environment in-
cluding: what unknown (i.e., in addition to known objects) dynamic objects exist in the
world, the number of objects on the table, the position of the objects on the table, the
position of objects on the ground that can obstruct a passage, the location of the garbage
can, and the state of the doors.
Name
aver.
min
max
planning/execution phases
62
56
73
action primitives per phase
5.14
1
27
action primitives per run
313
276
359
percepts
69
58
78
percepts per phase
1.13
1.04
1.20
planning CPU time per action
0.0062s
-
-
planning CPU time per phase
0.0317s
0.0005 s
0.1898 s
planning CPU time per run
1.9334s 1.5396s 2.2994s
execution time - action
3.363s
0.008s
34.17s
execution time - run
1056 s
897s
1259 s
(a) TASER picks up a garbage can.
(b) Results for the experiments with the physical service robot
TASE R.
Fig. 5.
We performed five runs with varying situations (e.g., different position of objects).
For all runs, the robot successfully performed the given task. Fig. 5(b) shows additional
results. On average, ACogControl divided the overall task of cleaning the table and
bringing the garbage out into 62 planning and execution phases, executed 313 action
primitives for an experiment run, and planned 5.14 steps (i.e., action primitives) ahead.
For the complete execution of the given tasks, 1.9334 seconds CPU time on average
is used for planning and reasoning. This is very low compared to the mean execution
time, which is 3.363 seconds for an action primitive and 1056 seconds for a complete
experiment run. Thus, the ratio of time used for planning and reasoning to the overall
execution time is very low at
1
.
9334 s
1
.
9334 s+1056 s
=0
.
0019
.
