Shepherding Via Deformable Shapes

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 7

Shepherding via Deformable Shapes

Joseph F. Harrison Christopher Vo Jyh-Ming Lien

Abstract— In this paper, we present a new motion planning


strategy for shepherding in environments containing obstacles.
This instance of the group motion control problem is applicable
to a wide variety of real life scenarios, such as animal herding,
civil crowd control, and oil-spill cleanup. The problem is
challenging in terms of scalability and robustness because it
is dynamic, highly underactuated, and involves multi-agent
coordination.
Our previous work showed that high-level probabilistic
motion planning algorithms combined with simple shepherd-
ing behaviors can be beneficial in situations where low-level
behaviors alone are insufficient.
(a) (b)
In this paper, we present a hybrid behavior called D EFORM,
in which the shepherds view the flock as an abstracted de-
formable shape. We show that our method is more robust than
our previous approach and that it scales more effectively to
larger teams of shepherds and larger flocks.

I. I NTRODUCTION
Group motion control is the problem of moving a group
of agents in coordination. An instance of this problem is
shepherding, where the objective is to herd a group of agents
(e.g. a flock of sheep, crowd of people, etc.) using one or
more “shepherd” agents (e.g. shepherds, riot police, etc.). (c) (d)
The objective of shepherding is typically to guide the group
to a goal, though other variants exist (e.g. escort or protection Fig. 1. Snap shots (a-d) of 15 shepherds successfully herding a
- a “parent” tries to maintain separation between their “child” large flock with 200 members exhibiting simple flocking behaviors
from one plaza to another in a street like environment.
and “stranger” [1]). A solution to the shepherding problem is
usually a sequence of movements the shepherd can perform
to guide the group to the goal. This sequence may be the for several reasons. First, it more accurately represents the
output of a high-level motion planner, the result of agent- contour of the flock than simpler representations such as
based behaviors working independently or in coordination to spheres or bounding boxes. Second, it favors areas of the
achieve high-level commands, or some combination thereof. search space where there is enough clearance for the blobs
Our recent work [2] showed that high-level planning to pass through.
combined with low-level shepherding behaviors can be an Our experiments show that D EFORM broadly outperforms
effective approach to guiding flocks amongst certain kinds our previous method (M AGB) in terms of robustness (higher
of obstacles. However, the results were inconsistent, and success rate), and scalability (effective for larger flocks and
on environments with narrow corridors and multiple paths larger teams of shepherds).
to the goal, our simplest method, “Medial Axis Graph-
Based” (M AGB) outperformed the planning-based methods II. R ELATED W ORK
when given a fixed budget of simulation steps. However, The group motion control problem is one that merges
M AGB does not scale well to larger flocks or larger teams the simulation and modeling of multi-agent dynamics with
of shepherds. cooperation, planning, and control. There is a large body
Main results This paper introduces a new approach to of work on the dynamics and modeling of human crowds,
shepherding called D EFORM, which is a hybrid of behavior- but there is still very little work on swarm motion control
based shepherd motions and low-level planning. The method via agent interactions, and no work has been proposed to
relies on an abstraction of the flock as a discretized de- systematically study group motion control as a whole.
formable shape or “blob”. We have chosen this abstraction Works that explore the crowd control problem have used
a variety of simple approaches such as introducing new
All authors are with Department of Computer Science, George Mason
University, 4400 University Drive MSN 4A5, Fairfax, VA 22030 USA, objects or agents into the environment. For example, the
{jharri1, cvo1, jmlien}@cs.gmu.edu effect of adding barriers into disaster scenario environments
was studied in [3]. Other works have attempted to model the Let the composite states of C at the time step t be SC (t) and
effect of adding agents with various social forces. The effect let those of G be SG (t). Therefore, a group control problem
of adding robots with attractive social forces was examined with n flock members and m shepherds in a 2-d workspace
in [4], and crowd dynamics in the presence of “leader” will have a state space in 4(n + m) dimensions.
individuals was modeled in [5]. Each flock member g ∈ G will be repelled by each visible
There are several works in robotics and computer ani- shepherd c ∈ C. We assume that the shepherd c knows about
mation related to modeling behaviors such as shepherding. g’s reactions. The flock G also exhibits certain behaviors
For example, Schultz et al. [6] applied a genetic algorithm when no shepherds are visible. In this paper, we will use
to learn rules for a shepherd robot. Its objective was to the flocking behavior [16]; however, other models such as
control the movement of other robots (sheep) that react by pedestrian dynamics [17] can also be used. In fact, the
moving away from the shepherd. Vaughan et al. [7] simulated proposed methods will be independent of G’s behaviors. In
and constructed a robot that shepherds a flock of geese addition to the collision-free constraint, other (user-defined)
in a circular environment. In computer animation, Funge constraints may be added, such as constraints that prevent
et al. [8] simulated an interesting shepherding behavior in the flock from separating for too long or prevent merging
which a T. Rex chases raptors out of its territory. Potter different types of flocks. A state that does not violate any
et al. [9] studied a herding behavior using three shepherds constraints is called a feasible state.
and a single sheep in a simple environment. None of the We use the term target to denote any intermediate position
aforementioned methods have shown the ability to guide that the shepherd attempts to steer the flock towards, and we
flocks in environments with obstacles. use the term steering point to denote any position to which
Swarm control can also be viewed as a cooperative prob- the shepherd moves himself in order to influence the move-
lem. The survey from Parker [10] provides an overview of ment of the flock. A roadmap is an abstract representation
multi-robot systems. From the perspective of multi robot sys- of the feasible configuration space in a given environment,
tems, the task of crowd control requires inherent cooperation, given as a directed graph G = (V, E), where each node in
in which the success of a robot in the team depends on the V represents a valid configuration, and each directed edge
actions of other robots. Inherent tasks (such as crowd control) (p, q) ∈ E denotes that it is possible for the flock to travel
are distinguished from non-inherent tasks (such as covering) from configuration p to configuration q.
in that they cannot easily be decomposed into independent
sub-tasks and thus are generally more difficult. Multiple B. Problem Statement
robots may also move in formation [11] to accomplish a Finally, we define the group-control problem as follows:
given task. In our previous research on shepherding behaviors Given SC (0) and PG (0), find a sequence of feasible states
[12], we observe that formations can be used effectively to SC (t) for 0 < t ≤ 1 so that all PG (t) are feasible and
control the motion of the group. Similar observations have PG (1) ∈ GS, where GS is a set of user specified goal states.
also been found in sociological studies of crowd control [13].
Recently, Shell and Matarić [14] have used multiple robots IV. OVERVIEW OF O UR M ETHOD
to deploy and assist with the evacuation of pedestrians. In this section, we will talk about the main framework
Effective solutions to the shepherding problem might also of the proposed method. This framework can be realized in
be applicable in other areas. The deformable shapes used in many different ways. An implementation of this framework
this paper have dynamic properties similar to oil spills [15], is discussed in Section V.
and this approach may be useful for developing effective
containment and cleanup strategies. A. Group Control as Deformable Shape Manipulation
Other Recent Work (note: integrate these into the other In order to control the flock’s motion, the shepherd’s
sections) Yeh et al. developed composite agents [1] which internal models should comprise the following components:
can exhibit many different behaviors including guidance, for (1) the shepherd’s behavior model, (2) the environment
which they use proxy agents as temporary obstacles. model, (3) the flock model, and (4) a model to represent
the other shepherds. Later, we will use these models as our
III. P RELIMINARIES
basis to define shepherd’s motion. Details of each model are
A. Group Control Problem described below.
In this section, we define some of the terms and concepts Behavior model The shepherd’s behavior is similar to
that are used later in the paper. that of a single- or multiple-arm robot performing object
In a group control problem, there are two types of agents: manipulation. To manipulate an object, a robot arm must
shepherds C and a flock G. A shepherd is an external agent alternate between two basic behaviors: “Transit” (reaching
that influences the movement of the flock. A flock is a and grasping the object) and “Transfer” (transporting the
collection of agents that try to steer away from the shepherd. grasped object) [18], [19]. The shepherd has two analogous
The shepherd’s task is to steer the flock to desired locations. behaviors: “Approach” (transit) and “Steer” (transfer). In
All agents have the shape of 2-d discs (of varying radii). approaching behavior, the shepherds attempt to get close
The state of an agent A (a shepherd or a flock member) is to the flock without disturbing it. In steering behavior, the
represented by the position and velocity of A, i.e., (xA , vA ). shepherd attempts to move the flock.
obstacle

(a) (b) (c)

Fig. 2. (a) An environment with a global roadmap, two obstacles, a shepherd (larger disc) and two sub-flocks. Each flock member is shown as a small
disc enclosed in a dashed circle, which represents its visibility. (b) The shepherd’s flock model is shown as two α-shapes for the two sub-flocks. (c) A
local roadmap is built around the flock nearest the shepherd.

Environment model All shepherds have access to a global problem. Given a desired shape, the shepherds’ task is to
roadmap representing the workspace. An example of such a find and position themselves at the contact points near the
roadmap is shown in Fig. 2(a). deformable object. A solution to the problem is likewise
Flock model It is impractical to handle each flock member reduced to determining a sequence of target shapes that will
individually, even when controlling a small flock. We believe eventually lead the flock to the final goal. D EFORM, shown
that the shepherd should have a high level representation in Algorithm 1, outlines the steps necessary to achieve the
of the flock to more efficiently guide the flock. In most goal.
existing methods, a bounding circle of the flock is used to
model the flock. This model works fine when the workspace Algorithm 1 D EFORM (G, C, g)
is relatively sparse. However, in environments with many Input: G (flock), C (shepherds), and g (goal)
obstacles or narrow corridors or when the flock size is large, while g is not reached do
a bounding circle is excessively restrictive. Compute the contour polygon PG (t) of G at time t
In the proposed framework, the shepherd represents the Determine the next target polygon PT (t + δt)
flock as a deformable object which can also split and Find ssteer ∈ SC to morph PG to PT
merge. During steering, the shepherd acts as a “deformer”, Move C to ssteer
continuously reshaping the contour of the flock to some
target shape. For example, the shepherds can represent the More precisely, D EFORM regards the flock as a deformable
contour of the flock using α-shape [20], [21], shown in object with an area conservative constraint. That is, D E -
Fig. 2(b), which is updated as the flock moves. To prevent the FORM will find a sequence of continuously deforming poly-
shepherd from disturbing the flock unnecessarily, the value gons PG (t), t ∈ [0, . . . , 1], so that PG (1) is near a goal. For
of α can be determined from the flock’s sensing range. In this each PG (t), the shepherds can tightly pack all flock members
paper, the shepherds represent the flock using “pixel blobs” in PG (t). Because D EFORM may not consider shepherd
(see Section V for details) to represent the flock’s deforming positions when building PG (t), additional soft constraints on
contour. PG (t) should be imposed to increase the chance of finding a
Shepherd model A shepherd needs to have a model to successful control plan. For example, the shepherds should
represent other shepherds. This model affects how we solve keep each polygon PG (t) as “fat” as possible and keep the
the the task allocation problem that assigns steering positions medial or the principal axis of PG (t) close to the edges of the
to shepherds. In this framework, we use a centralized system global roadmap. The fatness increases the controllability of
that knows the locations of all the shepherds. the flock and the centeredness reserves room for maneuver.
This is analogous to a large robot arm with many fingers There are many ways to compute the target polygons PT (t).
(i.e. shepherds) manipulating a deformable object (i.e. flock). For example, we can grow a search tree along the edges
All these fingers are connected to and only communicate with of the global (workspace) roadmap. Alternatively, we can
a central command center, which coordinates the shepherds. simply extract a path with maximal clearance and build a
Therefore, we reduce the task assignment problem to a sequence of area-conserving polygons along the path. Details
bipartite matching problem between the steering points and are discussed in the next section.
the shepherds. Then, the edge weight of the bipartite graph is Next, D EFORM computes the movements for the shep-
the geodesic distance (estimated using the global roadmap) herds so that the flock can assume the shape of polygon
between a shepherd and a steering point. PG (t) at each time step t. More specifically, given the
desired shape PT (t + 4t) in the next time step and the
B. Shepherding current flock model PG (t), the behavior will determine the
By representing the flock as a deformable object, the shep- necessary deformation and transformation to morph PG (t)
herding problem becomes a deformable object manipulation to PT (t + 4t) using, for example, the Iterative Closest Point
(ICP) algorithm [22] when α-shape is used to compute the B. Target Blobs
contour. In this paper, we use boolean operators between Given the current flock blob, the target shape is defined
pixel blobs to determine the necessary deformation. Finally, as another pixel blob, called a target blob, that includes all
D EFORM calculates a steering state ssteer ∈ SC for the flock members at the next time step. It is straightforward
shepherds (recall that SC is the state space of the shepherds). to compute the target blob. First, we pick a pixel that is
occupied by the closest flock member to the final goal
a potential steering point position. The closeness to the goal is measured by geodesic
distance. Once the center pixel o is determined, we grow the
PG − PT blob using free pixels, starting with the 8-connected set and
expanding in concentric rings around o. We stop when the
blob contains the same number of pixels as the number of
flock members. This ensures that the blob is large enough to
contain all flock members and also conserve the area of the
target blobs throughout the entire control behavior.
PT We say a flock configuration conforms to a target blob if
the target blob includes all flock members.
(a) (b)
Note that the target blob could be grown in many different
Fig. 3. (a) Target shape, current flock blob, and a potential steering point. ways. For example, one could make it grow away from
(b) Finding steering points using a radial partitioning. obstacles or toward the medial axis. The version we used
in this paper makes the blob grow as fat as possible.

C. Shepherds’ Steering Points


V. I MPLEMENTATION D ETAIL
Given a flock blob PG and a target blob PT , we compute
In this section, we flesh out the framework discussed in the set of steering points ssteer . To compute ssteer , we first
the previous sections. In particular, we discuss in detail how compute the differences between PG and PT . This is simply
a flock is represented as a deformable shape, how the target done by performing the boolean difference PG − PT pixel
shape is determined, and how the steering points are chosen. by pixel.
An important feature of our framework is that it can be Next, we determine a set of potential steering points,
implemented in many different ways. While using α-shapes s0steer . We say a pixel p is in s0steer if (1) p is neighboring
can provide high accuracy in flock contour representation, in (using 8-connectivity) to a pixel c ∈ PG − PT , (2) p 6∈
order to efficiently create the flock model, a new algorithm to (PG ∪ PT ), and (3) p is on the opposite side of the pixel in
efficiently update the α-shapes [23], [24], [25] is required. PT closest to c. An example of p is shown in Fig. 3(a).
This method will need to exploit the temporal and spatial Once the points in s0steer are identified, the last step is
coherence of the flock’s movement. to choose n points from s0steer , where n is the number of
To avoid this difficulty, we discritize the workspace into a shepherds. If the there are fewer than n points in s0steer ,
regular grid and represent the contour of the flock using pixel we are finished. In this case, some shepherds may not have
blobs (described below). Each cell in the grid is classified as steering points assigned and will stay stationary.
a free cell if a flock member in the cell is free of collision If there are more points in s0steer than there are shepherds,
from the workspace obstacle. Otherwise, the cell is marked then we partition s0steer into n pie wedges whose apex is at
as an in-collision cell. The size of the cell in the grid is the center of PT (the flock member closest to the global
defined by the geometric size of an individual flock member. goal). In each pie wedge, the point in s0steer farthest from
Therefore, each cell can only be occupied by a single flock the global goal is selected as a steering point (see an example
member. This approximation is very easy to implement and in Fig. 3(b)).
is very efficient (linear to the flock size). As we will see After ssteer is computed, the steering points are assigned
later in our experiments, this representation is also accurate to the shepherds by forming a bipartite graph whose nodes
enough to model and handle large flocks. are the points in ssteer and the shepherd positions, and
whose edge weights are geodesic distances between them.
The steering point assignment is then solved using a bipartite
A. Flock Blobs matching algorithm.
The pixel blob of a single member g of a flock is simply Screen shots our simulation running D EFORM can be
the set of pixels that are within the distance r − of g, where found in Fig. 4.
r is the sensor range of g and  is an arbitrarily small number.
VI. B EHAVIOR - BASED M OTION P LANNERS
The pixel blob of a flock (called flock blob) is therefore the
union of all pixel blobs of its members. The flock blob may D EFORM can also be used as a local planner in higher-
have several connected components if the flock has separated level planners such as PRM, RRT and EST. Based on our
into sub-flocks. results here and in [2], planning is not generally helpful.
potential steering points PG − PT PT

flock
shepherds

(a) “empty” (b) “s” (c) “spiral”

Fig. 4. Screen shots of the proposed shepherding behavior in action.

(d) “broken-t” (e) “pillars” (f) “env-6”


However, theoretically speaking, the proposed control be-
havior may not always successfully control the flock. For Fig. 5. Environments used in our experiments.
example, in some rare cases, the deformable polygon may
block paths needed by shepherds to reach their steering
points. For the sake of completeness, we will discuss two it can be easily adapted to the proposed method. Each node
high-level planners: RRT and META - GRAPH integrated with in a meta graph represents a meta configuration. A meta
the D EFORM. configuration defines a set of group configurations (excluding
the shepherd positions) that are conforming to the properties
A. Distance Metrics
of the node. More specifically, each meta configuration C is
In both planners, distance metrics are important for esti- as a pixel blob that has twice as many pixels as the number of
mating the the probability of successfully herding from one flock members. Each meta configuration intuitively defines
flock state to another flock state. Using Euclidean distance is a set of conforming flocks. We say a flock is conforming to
usually not enough to reflect this probability. In addition, we a meta configuration C if x% flock members overlap with
use the idea of deformation energy which defines the amount the pixels of C. In our implementation, we let x = 85. A
of effort needed to deform one polygon (representing the meta configuration is generated in the same way as the target
start state) to another polygon (representing the goal state). blob is generated (in Section V). We first pick a random
Again, deformation energy can be implemented in many point p in workspace and use p two grow a blob level by
ways depending on the representation of the deformable level. These meta configurations are then connected to their
object. By representing the polygon as a pixel blob, we k-nearest neighbors form a meta graph. In a similar fashion
have the advantage of efficiently computing the overlay of to lazy PRM approaches [19], paths from the meta graph are
two deformable polygons. More specifically, given two pixel extracted and evaluated using the proposed control behavior
blobs, P and Q, we translate P so that their centers coincide, until a path is found, or until no path can be found to connect
then count the number of pixels in the boolean difference the start and the goal configurations in the meta graph.
P −Q. This value approximates the number of pushes needed
to deform P to Q. The final distance between P and Q VII. E XPERIMENTAL R ESULTS
is simply a weighted sum of the Euclidean distance be- We evaluated our new D EFORM behavior against the
tween their centers and the deformation energy estimated by simpler M AGB behavior across 6 different test environments,
boolean difference P −Q. In our implementation, we use the shown in Figure 5. The next sections explain our experimen-
same weights for both Euclidean distance and deformation tal method. For this paper, each experimental sample consists
energy. of 30 runs, and claims of statistical significance are based
B. RRT on 95% confidence.
Our RRT-based planner constructs the roadmap G by A. Scalability
repeatedly attempting to extend G using the proposed control
behavior towards new randomly generated configurations in We presume that increasing the size of the flock should
W . For RRT, the procedure S ELECT-I NTERMEDIATE G OAL impact both the effectiveness and performance of the shep-
chooses a random configuration g in the workspace, and herding algorithm, and that more shepherds will be needed
the procedure S ELECT-N ODE -T O -E XPAND chooses the node to move larger flocks to the goals in a timely manner. The
already in G that is closest to g. scalability of these behaviors is defined by the ability for
them to effectively herd increasingly large flocks to the goal
C. Meta Graph within the given simulation time budget.
Meta graph is a PRM-based method proposed in our recent To test our hypotheses, we ran experiments with varying
work [2]. Although originally designed for directed circles, number of shepherds (1, 2, 3, and 4) and varying flock sizes
randomness. In these figures, randomness is shown as a
1.2 DEFORM
MAGB ratio from 0.0 to 1.0 where 0.0 represents fully determin-
stic behavior and 1.0 represents fully random behavior. In
1
general, D EFORM performs significantly better than M AGB
0.8 with increasing flock randomness. An interesting effect is
Success Rate

shown in Figure 7, where there is a bump in performance


0.6
with increasing flock randomness for the M AGB behavior.
0.4 We believe that this is because the random oscillation of
the flock members helps the shepherds push them through
0.2 narrow corridors in the broken-t environment in similiar way
0 to salt through the openings of a salt-shaker.

-0.2 1.2 DEFORM


1/05 2/10 4/20 6/30 8/40 MAGB
Shepherd/Flock Ratio 1

Fig. 6. Success rate vs. Shepherd/Flock Ratio on s Environment. 0.8

Success Rate
0.6
(5, 10, 15, 20, and 25) on each of the 6 test environments. 0.4
We computed success rates as the proportions of sample
runs where the algorithm successfully moved the flock to 0.2
the goal. Table I shows example results for the “broken- 0
t” environment. For small flock sizes, D EFORM and M AGB
perform similarly. However, as the size of the flock increases, -0.2
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
M AGB shows significant decay in performance - it is clear Flock Randomness
that more shepherds are needed to control larger flocks using
the M AGB behavior. On the other hand, D EFORM manages Fig. 7. Success Rate vs. Randomness Ratio on the broken-t Environment.
to achieve excellent results throughout (above 90% sucess
rate) with no negative trend in performance up to 25 flock
members.1 1.2 DEFORM
MAGB
We also ran experiments to test larger shepherd and flock
1
sizes on the “s” environment. The results, shown in Fig. 6,
show that D EFORM once again outperforms M AGB across 0.8
Success Rate

the board, but especially with larger flock sizes.


0.6
TABLE I
0.4
S UCCESS R ATES FOR D EFORM AND M AGB WITH VARYING S HEPHERD
AND F LOCK S IZES (B ROKEN -T) 0.2

D EFORM M AGB 0
# Shepherds # Shepherds
1 2 3 4 1 2 3 4 -0.2
5 0.93 1.00 0.90 0.83 0.90 1.00 0.93 0.97 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
Flock Size

10 1.00 0.93 0.93 0.93 0.56 0.97 1.00 0.97 Flock Randomness
15 1.00 1.00 1.00 0.93 0.47 0.80 0.77 0.83
20 1.00 0.97 1.00 0.90 0.07 0.67 0.43 0.83 Fig. 8. Success Rate vs. Randomness Ratio on the s Environment.
25 1.00 1.00 0.97 0.97 0.00 0.40 0.33 0.43

C. Does High-Level Planning Benefit from D EFORM?


B. Robustness
Our previous work compared the effectiveness of a simple
We define robustness as the ability for the shepherds to
behavior-based shepherding algorithm to algorithms that
control the flock in spite of unpredictable flock behavior.
combined low-level behaviors with high-level planning. We
We modeled the unpredictable behavior by linearly mixing
found that in many cases, the planning methods we applied
each flock agent’s behavior with a random vector. We control
did not yield significant enough improvements to justify the
this randomness by increasing and decreasing the magnitude
additional computational expense. Indeed, our tests showed
of the random vector. Figures 7 and 8 illustrate the results
that with a fixed budget of simulation steps, planning-based
(with 95% confidence bars) of success rate versus increasing
approaches using these techniques were often unable to
1 We intend to extend these experiments to higher flock sizes for the final even match the performance of our behavior-only method.
release of this paper. In almost all environments, planners using the D EFORM
behavior performed significantly better than planners using [6] A. C. Schultz, J. J. Grefenstette, and W. Adams, “Robo-shepherd:
the M AGB behavior. We are still studying how to make use Learning complex robotic behaviors,” in In Robotics and Manufactur-
ing: Recent Trends in Research and Applications, Volume 6. ASME
of these shepherding behaviors with higher-level planners for Press, 1996, pp. 763–768.
better performance in difficult environments. [7] R. T. Vaughan, N. Sumpter, J. Henderson, A. Frost, and S. Cameron,
“Experiments in automatic flock control,” J. Robot. and Autonom. Sys.,
TABLE II vol. 31, pp. 109–117, 2000.
[8] J. Funge, X. Tu, and D. Terzopoulos, “Cognitive modeling: Knowl-
S UCCESS R ATES FOR D EFORM AND M AGB W ITH VARIOUS
edge, reasoning and planning for intelligent characters,” in Computer
H IGH -L EVEL P LANNERS AND E NVIRONMENTS Graphics, 1999, pp. 29–38.
[9] M. A. Potter, L. Meeden, and A. C. Schultz, “Heterogeneity in the
D EFORM M AGB coevolved behaviors of mobile robots: The emergence of specialists,”
Planner Planner in IJCAI, December 2001, pp. 1337–1343.
None Graph Tree None Graph Tree [10] L. E. Parker, “Current research in multi-robot systems,” Journal of
broken-t 0.97 0.13 1.00 0.86 0.00 0.07 Artificial Life and Robotics, vol. 7, 2003.
Environment

empty 1.00 0.00 0.77 1.00 0.00 0.33 [11] T. Balch and R. Arkin, “Behavior-based formation control for multi-
env6 1.00 0.30 0.97 0.10 0.00 0.00 robot teams,” IEEE Trans. Robot. Automat., vol. 14, no. 6, pp. 926–
pillars 0.97 0.96 1.00 0.93 0.00 0.00 939, 1998.
s 0.97 1.00 0.83 0.70 0.00 0.00 [12] J.-M. Lien, O. B. Bayazit, R.-T. Sowell, S. Rodriguez, and N. M.
spiral 0.83 0.70 0.53 0.00 0.00 0.00 Amato, “Shepherding behaviors,” in Proc. IEEE Int. Conf. Robot.
Autom. (ICRA), April 2004, pp. 4159–4164.
[13] R. Applegate, Riot control: materiel and techniques. Stackpole Books,
VIII. C ONCLUSION AND F UTURE W ORK 1969.
[14] D. A. Shell and M. J. Matarić, “Directional audio beacon deployment:
This paper introduces and applies a new type of abstraction an assistive multi-robot application,” in Proc. IEEE Int. Conf. Robot.
to the problem of shepherding. By representing the flock Autom. (ICRA), 2004, pp. 2588–2594.
[15] X. Chao, N. J. Shankar, and H. F. Cheong, “Two- and
configuration as a discretized deformable object, we show three-dimensional oil spill model for coastal waters,” Ocean
that shepherding can be done more effectively than our pre- Engineering, vol. 28, no. 12, pp. 1557 – 1573, 2001. [On-
vious best method [2] in terms of handling larger flock sizes line]. Available: http://www.sciencedirect.com/science/article/B6V4F-
43MKKTY-3/2/26621349ce0865179e08c508afb58f69
and unpredictable flock behavior. This algorithm also has the [16] C. W. Reynolds, “Flocks, herds, and schools: A distributed behaviroal
virtue of needing less information about the environment. model,” in Computer Graphics, 1987, pp. 25–34.
Whereas our previous method relied on the medial-axis of [17] P. Kachroo, S. J. Al-nasur, S. A. Wadoo, and A. Shende, Pedestrian
Dynamics: Feedback Control of Crowd Evacuation. Springer, 2008.
the map, this new algorithm needs only an occupancy grid [18] Y. Koga, K. Kondo, J. Kuffner, and J. Latombe, “Planning motions
indicating open and closed tiles. This makes it significantly with intentions,” in Proc. ACM SIGGRAPH, 1995, pp. 395–408.
more applicable to robotics scenarios where a priori map [19] C. L. Nielsen and L. E. Kavraki, “A two level fuzzy PRM for manip-
ulation planning,” IEEE/RSJ International Conference on Intelligent
information is not available. Robotics and Systems, pp. 1716–1722, 2000.
In the future, we’d like to try growing target blobs that [20] H. Edelsbrunner, D. G. Kirkpatrick, and R. Seidel, “On the shape of
favor metrics besides just “fattest”, such as growing away a set of points in the plane,” IEEE Trans. Inform. Theory, vol. IT-29,
pp. 551–559, 1983.
from obstacles or towards planned waypoints. We are also [21] H. Edelsbrunner and E. P. Mücke, “Three-dimensional alpha shapes,”
further exploring the interaction between high level planning ACM Trans. Graph., vol. 13, no. 1, pp. 43–72, Jan. 1994. [Online].
and these behaviors. Available: ftp://cs.uiuc.edu/pub/edels/geometry/shapes-94.tar.Z
[22] P. J. Besl and N. D. McKay, “A method for registration of 3-d shapes,”
IEEE Trans. Pattern Anal. Mach. Intell., vol. 14, no. 2, pp. 239–256,
R EFERENCES 1992.
[1] H. Yeh, S. Curtis, S. Patil, J. van den Berg, D. Manocha, and M. Lin, [23] H. Cheng, H. Edelsbrunner, and P. Fu, “Shape space from defor-
“Composite agents,” in Proceedings of Eurographics / ACM SIG- mation,” in PG ’98: Proceedings of the 6th Pacific Conference on
GRAPH Symposium on Computer Animation, M. Gross and D. James, Computer Graphics and Applications. Washington, DC, USA: IEEE
Eds., 2008. Computer Society, 1998, p. 104.
[2] C. Vo, J. F. Harrison, and J.-M. Lien, “Behavior-based motion planning [24] S.-W. Cheng, H. Edelsbrunner, P. Fu, and K.-P. Lam, “Design and
for group control,” in Proceedings of the IEEE International Confer- analysis of planar shape deformation,” Comput. Geom. Theory Appl.,
ence on Intelligent Robots and Systems (IROS), St. Louis Missouri. To vol. 19, pp. 205–218, 2001.
Appear., 2009. [25] H.-L. Cheng, T. K. Dey, H. Edelsbrunner, and J. Sullivan, “Dynamic
[3] M. Brenner, N. Wijermans, T. Nussle, and B. de Boer, “Simulating skin triangulation,” in SODA ’01: Proceedings of the twelfth annual
and controlling civilian crowds in robocup rescue,” in Proceedings of ACM-SIAM symposium on Discrete algorithms. Philadelphia, PA,
RoboCup 2005: Robot Soccer World Cup IX, 2005. USA: Society for Industrial and Applied Mathematics, 2001, pp. 47–
[4] J. Kirkland and A. Maciejewski, “A simulation of attempts to influence 56.
crowd dynamics,” in IEEE International Conference on Systems, Man
and Cybernetics, vol. 5, 2003, pp. 4328–4333.
[5] F. Aubé and R. Shield, “Modeling the effect of leadership on crowd
flow dynamics.” in ACRI, ser. Lecture Notes in Computer Science,
P. M. A. Sloot, B. Chopard, and A. G. Hoekstra, Eds., vol. 3305.
Springer, 2004, pp. 601–621.

You might also like