Professional Documents
Culture Documents
Department of Computing, The Hong Kong Polytechnic University, Hung Hom, Hong Kong
E-MAIL: {csmwang, csnkliu}@comp.polyu.edu.hk
Abstract:
This paper proposes a new method, minimum risk 2. The path searching behavior
approach, to address the local path planning to escape from
local minimum during goal-oriented robot navigation in The Path-Searching (PS) behavior is one application of
unknown environments. This approach is theoretically proved
our memory grid technique. The memory grid is a new type
to guarantee global convergence even in the long-wall,
unstructured, cluttered, maze-like, and modified environments. of map for modeling the robot environment. Memory grid
The approach adopts a strategy of multi-behavior differs from other maps (e.g. occupancy grid, certainty grid,
coordination, in which a novel path-searching behavior is connectivity graph [8, 9]) chiefly in that it makes use of an
developed to recommend the regional direction with minimum obstacle and trajectory memory dot model. The memory
risk. The paper provides a fuzzy logic framework to grid uses the obstacle memory dots to quantify the fuzzy
implement the behavior design and coordination. It is verified possibility that represents the uncertainty of obstacles
by the simulated and real world tests. detected by sonar sensors, while it uses the trajectory
memory dots to save the trajectory traversed by robot. The
Keywords: longer the robot takes to traverse a grid region, the greater
Fuzzy control; intelligent control systems; mobile robot the number of trajectory memory dots saved in this grid
motion planning; local path planning
(see Fig.6(9)). Therefore, the regional direction with the
minimum number of obstacle and trajectory memory dots
1. Introduction can represent the regional direction with the minimum risk.
This direction is the good choice for the robot in trying to
We call it “blind goal-oriented navigation” that the avoid a previous trajectory, and consequently escape from
robot is required to autonomously reach a desired goal but local minimum and reach the goal. PS behavior does this
it does not have a priori known environmental knowledge. job to recommend the direction that entails minimum risk.
The related methods [1-7] can be categorized as two types:
boundary following, and virtual subgoal. Unfortunately, 2.1. Regional Risk Index
they have difficulties to guarantee the global convergence
in the complex environments since they may get trapped in
Four features are extracted from memory grid data, in
the local minimum (or dead end). This paper proposes a
which iteration risk α and collision risk β are used to
new method, minimum risk approach, to address the local
path planning to handle the local minimum during the blind infer the Risk Index for the fuzzy navigational rules of PS
goal-oriented robot navigation. This approach adopts a behavior, and the trajectory dot intensity κ and obstacle
strategy of multi-behavior coordination, in which a novel dot intensity τ are used in combination with fuzzy logic to
Path-Searching (PS) behavior is developed to recommend calculate the weight of PS behavior. The magnitude of
the regional direction that is offering minimum risk. The iteration risk α is converted into three linguistic fuzzy
final command output is obtained by coordinating three sets {LOW, MEDIUM, HIGH}, with the membership functions
behaviors: PS, Obstacle-Avoidance (OA), and shown in Fig. 1(a). The other three are similar.
Goal-Seeking (GS). The paper provides a fuzzy logic The Fuzzy Rule-Based Risk Index combines the two
framework to implement the behavior design and regional risk parameters into a single indicator of safety of
coordination. Thus errors due to sensor noise and traversal of the region by the mobile robot. The Risk Index
self-localization are effectively handled by our navigation r is represented by three linguistic fuzzy sets {DANGEROUS,
system. UNCERTAIN, SAFE}, with the membership functions shown
µ (α ) µ (r )
low medium high dangerous uncertain safe
1 1
α r
0 α1 α2 α3 α4 α5 0 r1 r2 r3 r4 r5
(a) (b)
Fig. 1. Membership functions (a) for iteration risk. (b) for Risk Index
814
Proceedings of the Fourth International Conference on Machine Learning and Cybernetics, Guangzhou, 18-21 August 2005
important note, a turn maneuver is not initiated when the turn, but the left and right sectors have the same obstacle
three sectors have the same dangerous risk indices as distance, then the recommended turn angle is GOAL, where
shown in the (1,1) element of the top layer in Table II. The GOAL implies that the recommended turn angle should be
turn rule does not at this stage force the robot to arbitrarily toward the direction close to the goal location. This is
choose between left and right, but maintains the turn angle similar to the turn rules for PS behavior. One last important
at zero. The final selection will be made by a complement note: when the three sectors have the same VERYNEAR
algorithm, Minimum Collision & Iteration (MCI). The MCI obstacle distance as shown in the (3,3) element of the top
algorithm is exact, not fuzzy. layer in Table IV(a), a large left turn (PB) angle is
recommended. This turn rule enables the robot to escape
2.3. Weight rules from its current embarrassed situation.
d right
The weighting factor wps represents the strength by far near verynear
d front
d left far
which the PS behavior recommendation is taken into GOAL PS PB
verynear
account to calculate the final motion command. The weight near
NS GOAL PS
of PS behavior is represented by three linguistic fuzzy sets verynear
NB NS PB
{SMALL, MEDIUM, LARGE}, and is derived directly from d right
both the trajectory dot intensity κ and obstacle dot
far near verynear
dleft far d front
GOAL PS PS
intensity τ of a square region, using the rule sets in Table near
near
NS ZE PS
III. verynear
Table III. Weight rules of the PS behavior NS NS ZE
(a)
d right
3. The obstacle avoidance behavior far near verynear
dleft far d front
large large large
The local Obstacle-Avoidance (OA) behavior is near
verynear
large large large
actually a sensor-based behavior which implements a verynear
large large large
control strategy based on external sensing. OA behavior is d right
effective if obstacles are close. We adopt a fuzzy logic far near verynear
d front
controller for the implementation. The sonar readings of the dleft far
medium large large
near
robot are grouped into three sectors (left, front, right). For near
large large large
example, our robot has a ring of eight forward ultrasonic verynear
large large large
sonars that produce a set of obstacle distances {d0, d1, d2, d right
d3, d4, d5, d6, d7}. We obtain three groups of obstacle far near verynear
dleft far d front
distances by the following equations: dleft = min(d0, d1); small small large
far
dfront = min(d2, d3, d4, d5); dright = min(d6, d7). The near
small medium large
obstacle distance of each sector is represented by three verynear
large large large
linguistic fuzzy sets {VERYNEAR, NEAR, FAR}. The obstacle
(b)
avoidance navigation rules are discussed below. Table IV. (a) Turn rules for OA behavior. (b) Weight rules for OA behavior
The turn rules for the OA behavior are summarized in
Table IV(a). Observe that the rules exhibit such a behavior The robot speed v is based on the close obstacle
characteristic: if the obstacle distance in any sector is distance in the three sectors. The speed rules are as follows.
VERYNEAR, the robot should turn away to find a safer 1) IF dfront is VERYNEAR, THEN v is STOP.
direction. For instance, the (1,3) element of the bottom 2) IF dfront is NEAR, THEN v is SLOW.
layer in Table IV(a) can be written out as the rule: 3) IF dfront is FAR AND dleft is FAR AND dright is FAR, THEN v is
IF dfront is FAR AND dleft is FAR AND dright is VERYNEAR, THEN FAST.
4) IF dfront is FAR AND (dleft is VERYNEAR OR dleft is NEAR OR
θ oa is PS. dright is VERYNEAR OR dright is NEAR), THEN v is SLOW.
In addition, in Table IV(a), when the robot needs to The last rule enables the robot to decrease its speed
815
Proceedings of the Fourth International Conference on Machine Learning and Cybernetics, Guangzhou, 18-21 August 2005
when an obstacle is closing. Because in the environmental goal-oriented navigation task, this solution path must have
maps of the tests the exit location is very narrow and often the minimum obstacle and trajectory memory dots, i.e. the
situated in the lateral wall, a fast speed may cause the robot minimum risk. And the detected regions of the current robot
to miss the exit. position must have a regional direction that has a minimum
Like the weight of PS behavior, the weight of OA risk. The minimum risk approach can guarantee to finally
behavior woa is represented by three linguistic fuzzy sets recommend the region direction with minimum risk. Thus
{SMALL, MEDIUM, LARGE}, and is derived directly from the solution path must be found and the robot finally
obstacle distances in the three sectors. The weight rules for reaches the goal. That’s the global convergence. ■
the OA behavior are summarized in Table IV(b). Corollary 1: Even in the modified environment, the
minimum risk approach can guarantee global convergence
4. The goal seeking behavior if a solution path exists.
816
Proceedings of the Fourth International Conference on Machine Learning and Cybernetics, Guangzhou, 18-21 August 2005
-80
D Time
C
-100
0 50 100 150 200 250 300 350 400 450 500
(9) (10)
Fig. 6. Minimum risk approach to the long-wall environment with local minimum. (1-8) S is the start, T the goal target. OA is the red line “a”,
PS is the blue line “b”, GS is the black line “c”. (9) Underlying memory grid is shown by the spaced horizontal and vertical lines; The obstacle
memory dots are drawn as the black squares of different size. The larger is the size of the square, the higher the possibility of the obstacle is; The
trajectory memory dots are drawn as the red circles of different size. The larger is the circle, the more the number of trajectory dots is. (10) Turn
angles recommended by different behaviors. For the display, the GS turn angle is a half.
Consequently, the weight of PS behavior increases, and PS environment as shown in Fig.8(a), because the robot
behavior (line b) is effective in these cases (see encounters another local minimum at the location “b” and
(2)(3)(4)(5)(6) in Fig.6); When the front of the robot is “c” when it is working under the influence of previous
closing upon the obstacles, the weight of OA behavior (line virtual subgoal. Krishna and Kalra’s method [3] has a good
a) becomes large (see (2)(4)(6) in Fig.6); When the OA or result as shown in Fig.8(b). But it highly depends on the
PS behaviors are dominant, the GS behavior is suppressed landmark recognition and exact coordination localization.
and its weight is small (see (2)(3)(4)(5)(6) in Fig.6). When In addition, it is difficult to choose a correct direction to
the robot is far from the obstacles and is approaching the follow the wall boundary as seen in Fig.8(e). Maaref and
goal T at a normal speed, the weights of both OA and PS Barret’s method [4] fails to reach the goal in such large
behavior are small and only GS behavior is dominant (Fig.6 concave environment, because it detects the local minimum
(7)). Fig. 6(9) shows the underlying memory grid as well as using a restricted criterion that all sensors must give the
the control interface. small obstacle distances at the same time. Fig.8(c) shows
Observe that in Fig. 6 (8) and (10), the labels A, B, C, the result of our minimum risk approach. The robot exhibits
D, E, F, S, and T represent the robot locations. The turn the typical “trial-and-return” behavior phenomenon. This
angles recommended by OA and PS behavior are consistent property can help the robot to find the nearest exit to escape
during most of the entire task period. At the locations B, C, from the local minimum. It’s further verified in the Fig.8(f).
and D, the goal T is switched from the left of the robot to Fig.8(d) shows the result of Huang and Lee’s method [1].
the right, or from the right to the left. This is why the robot This method has a conservative leaving criterion that makes
leaves the wall at location D and turn toward location E. the robot traverse a longer path compared with other
Fig.7(a) and (b) show the results in the maze-like, methods that adopt the boundary-following strategy. More
unstructured and cluttered environments respectively. importantly, it is still difficult to choose the correct
We compare our minimum risk approach with other boundary-following direction. Similar problems occur on
related methods. The virtual target method [7] fails to reach the Distbug method [2] and Virtual-target-side method [5].
the goal in such concave and recursive U-shape Now we exhibit a real world test in a modified
817
Proceedings of the Fourth International Conference on Machine Learning and Cybernetics, Guangzhou, 18-21 August 2005
References
818