![](bg6.png)
fn
0.5 1.0 1.5 2.0 2.5 3.0
4
2
#
t
react
∆t
2∆t
Figure 5: Reaction time t
react
in [s] and one false-negative
(fn) of examples while facing different cracks.
1.0
0.5
0.0
0.0 0.25
0.5 0.75 S
haz
A
1.0
E
max
S
max
A
◦•
fp
◦
◦
◦
◦
◦
◦
◦
◦
◦
◦
⋆
Figure 6: Maximum evaluation and rating values on differ-
ent rough terrains, ⋆ marks tolerated false-positives (fp).
false-positives is important. Therefore, the reaction
on rough terrain with defects has been tested which
do not lead necessarily to a drop-off. Figure 6 shows
12 test runs with only one false-positive (black circle)
and two detections (E
max
= 1, marked with ⋆) which
are tolerated because of S
max
A
> S
haz
A
. In practice, the
evaluation system has to be trained once and can be
applied to similar situations and setups.
Each detection of safety-critical situations is use-
less without counteractive measures. So far, a re-
versed replay of the robot trajectory is implemented.
The responsible behavior has been embedded into the
control system and is stimulated, if the evaluation
function E reaches a value of 1. In this case, the cur-
rent driving operation is cancelled and the last com-
mands are countervailed. The idea is that the way was
not dangerous so far so the robot should driveback the
same trajectory until a safe position has been reached
and the adhesion system can recover.
6 CONCLUSIONS
This paper presented a risk prediction approach for
wall-climbing robots. Based on training data a ge-
netic algorithm is used to find suitable weights for a
general evaluation function which is used here to pre-
dict an upcoming drop-off. Experiments have proven
the functionality of the approach and the benefit for
robot safety. The next step is to adapt the prediction
system to be able to handle different situations (e.g.
driving up or down) which need to use differing sets
of weights since one-fit-all-weights do not exist.
REFERENCES
Castelnove, M., Arkin, R., and Collins, T. R. (2005). Re-
active speed control system based on terrain rough-
ness detection. In IEEE International Conference on
Robotics and Automation, volume 1, pages 891–896.
Gerdes, I., Klawonn, F., and Kruse, R. (2004). Evolution¨are
Algorithmen. Vieweg Verlag, Germany, 1. edition.
Kelly, A. and Stentz, A. (1998). Rough terrain autonomous
mobility - Part 1: A theoretical analysis of require-
ments. Autonomous Robots, 5(2):129–161.
Kim, D., Sun, J., Oh, S. M., Rehg, J. M., and Bobick, A. F.
(2006). Traversability classification using unsuper-
vised on-line visual learning for outdoor robot navi-
gation. In IEEE International Conference on Robotics
and Automation (ICRA) 2006, pages 518–525.
Luk, B. L., Cooke, D. S., and others (2001). Intelligent
Legged Climbing Service Robot For Remote Inspec-
tion And Maintenance In Hazardous Environments. In
8th Conference on Mechatronics and Machine Vision
in Practice, pages 252–256.
Proetzsch, M., Luksch, T., and Berns, K. (2010). Develop-
ment of complex robotic systems using the behavior-
based control architecture iB2C. Robotics and Au-
tonomous Systems, 58(1):46–67.
Schmidt, D. and Berns, K. (2011). Behavior-based adhesion
control system for safe adherence of wall-climbing
robots. In 14th International Conference on Climb-
ing and Walking Robots (CLAWAR), pages 857–864.
Schmidt, D., Hillenbrand, C., and Berns, K. (2011). Om-
nidirectional locomotion and traction control of the
wheel-driven, wall-climbing robot, Cromsci. Robot-
ica Journal, 29(7):991–1003.
Shang, J., Bridge, B., Sattar, T., Mondal, S., and Brenner,
A. (2008). Development of a climbing robot for in-
spection of long weld lines. Industrial Robot: An In-
ternational Journal, 35(3):217–223.
Spears, W. M. (1993). Crossover or mutation. In Foun-
dations of Genetic Algorithms, volume 2, pages 221–
237.
Stavens, D. and Thrun, S. (2006). A self-supervised terrain
roughness estimator for off-road autonomous driving.
In Conference on Uncertainty in Artificial Intelligence
(UAI), pages 469–476.
RiskPredictionofaBehavior-basedAdhesionControlNetworkforOnlineSafetyAnalysisofWall-climbingRobots
123