1

A Dynamic, Real-Time Testbed for Resource Management Technology

David Chelberg, Lonnie Welch, Cynthia Marling, Carl Bruggeman, Douglas Lawrence, David Matolak, Robert
Williams, Jae Lew, Arvind Lakshmikumar, Matthew Gillen, Qiang Zhou
(chelberg,welch)@ohiou.edu,(marling,bruggema)@p1.cs.ohiou.edu,
(dal,dmatolak,bobw,jlew,kla)@bobcat.ent.ohiou.edu,(mgillen,qzhou)@p1.cs.ohiou.edu
Ohio University
Barbara Pfarr
[email protected]
NASA Goddard Space Flight Center

Abstract—This paper describes a test-bed for technology ment model for agent based systems that have real-time
that unifies agent based computing and adaptive resource constraints. The model will be evaluated by applying it
management for dynamic real-time systems. We describe to two ground-based testbed prototypes of future NASA
a unified framework that combines a hybrid agent based systems: a satellite constellation command and control
architecture with explicit resource adapting mechanisms.
system, and a group of cooperating autonomous mobile
robots. The constellation command and control system
I. I NTRODUCTION is discussed in [17]. This paper represents the robot
Space Missions of the 21st Century will be character- testbed. The effectiveness of adaptive resource man-
ized by constellations of distributed spacecraft, miniatur- agement will be applied to a coopearting group of au-
ized sensors and satellites, autonomous mobile robots, tonomous mobile robots. Scenarios will involve the tasks
increased levels of automation and intelligent on-board of learning, planning, vision and navigation in unknown
processing. New technology will be infused to achieve environments. The tasks will have timing constraints as-
desired levels of autonomy and processing capability. sociated with them. A distributed set of computers will
Agent based computing offers the ability to decentral- be managed dynamically to allow the agents controlling
ize computing solutions by incorporating autonomy and the robots to make the best decisions possible under their
intelligence into cooperating, distributed applications. It resource and timing constraints.
provides an effective medium for expressing solutions to The robot agents will perform the RoboCup Chal-
problems that involve interaction with real-world envi- lenge. It represents a standard problem on fast-moving
ronments and allows modeling of the world state and multiple robots, which collaborate to solve dynamic
its dynamics. This model can be then used to deter- problems [6]. It is designed to meet the need of han-
mine how candidate actions affect the world, and how dling real world complexities. In RoboCup, robots com-
to choose the best from a set of actions. Most agent pete in a soccer game. RoboCup can be viewed as a
paradigms overlook real-time requirements and comput- multi-agent system, where software agents provide de-
ing resource constraints. For a real-time system, the right cision support for dispatching and engaging resources.
answer late is wrong. The effectiveness of agent comput- In such a system, knowledge, action and control are dis-
ing can be enhanced by combining it with system-level tributed where each agent may cooperate, compete or co-
resource allocation optimization approaches. Adaptive exist. Past approaches [14] have been monolithic in their
resource management approaches for real-time systems design, wherein an agent (an individual soccer player) is
focus on guaranteeing that real-time requirements are a local process or a group of processes on a single ma-
met with a particular set of computing resources, even chine. During a game of soccer, computing needs may
in dynamically changing environments. The authors are increase depending on the state of the game. If these
performing research which is unifying, for the first time, processes cannot migrate to a less loaded processor, a
the agent based computing paradigm with the theory single machine could get overloaded, causing an agent to
of adaptive resource management for dynamic real-time fail to complete its computations in the desired time. In
systems. This effort will produce a resource manage- general, without some means of dynamic resource man-

0-7695-0990-8/01/$10.00 (C) 2001 IEEE

Authorized licensed use limited to: UNIVERSITY OF SYDNEY. Downloaded on April 05,2010 at 23:37:28 EDT from IEEE Xplore. Restrictions apply.
 2

agement, there is an inefficient allocation of resources. SUN SUN Global Camera

HUB Workstation Workstation (Overhead)
Our approach to achieve this objective can be categorized
into two main areas, distributed agent based computing LAN
Access
Touch Sensors IR SENSORS KICKING
DEVICE

and resource management under dynamic, real-time con- HARD

DRIVE

straints. This paper is a description of the different as- Wireless Embedded MOTOR
LAN PC
pects involved in our RoboCup testbed. DRIVER

CAMERA 1 Frame MOTOR

Robots
Grabber Frame
Grabber
II. R EAL -T IME S YSTEMS Robots
CAMERA 2
MOTOR
Robots
POWER
A real-time system may be characterized as being SUPPLY
ROBOT

static or dynamic. A static system is one whose resource

requirements do not change with changes in its environ-
Fig. 1. Block diagram of the entire system
ment. A dynamic system [13] is one whose resource re-
quirements change unpredictably at runtime. This char-
acterization is based on the temporal properties and exe- provide the notion of large grain entities with end-to-end
cution behavior of the system. The majority of real-time timing constraints.
computing research has focused on systems whose re- To address these problems, we will use DynBench [18]
quirements are evaluated statically. However, RoboCup and RoboCup as benchmarks.
is a highly dynamic environment. This precludes the ac-
III. T ESTBED
curate characterization of the environment and its prop-
erties by static models. In such contexts, temporal and This section gives an overview of the physical and
execution characteristics can only be determined accu- software architecture of robots for the RoboCup chal-
rately by empirical observation or experience (i.e., a pos- lenge problem.
teriori). In most real-time computing models, the ex-
ecution time of a job is used to characterize workload A. Hardware Design
statically as an integer worst-case execution time [2], The physical robot (Figure 1) is made up of an omni-
[11], [5]. This is often difficult and sometimes impossi- directional base, a ball-dribbler and a kicker. The control
ble. The DeSiDeRaTa (Dynamic, Scalable, Dependable circuitry on board the robot includes a microprocessor,
Real-Time System)[12] middleware was designed to ad- memory and a wireless receiver. Our hardware system
dress these shortcomings. One of the fundamental inno- can be broadly classified into Mechanical components,
vations of DeSiDeRaTa is the dynamic path paradigm, Control circuits, Processor and communication devices.
that is employed for modeling and resource management The mechanical design of our robots was developed
of distributed, real-time systems. keeping in mind the agility and speed of the robot. The
To determine the practicality of the resource manage- robots must also be able to kick and control the ball
ment model, we need a benchmark suite and environ- accurately. To provide greater mobility, our robots are
ment for experimental assessment of adaptive QoS and equipped with an omnidirectional base. The main fea-
resource management software. The benchmark places tures of an omnidirectional system are:
load on computation and communication resources by  Ability to move in any direction
execution of software that performs control system func-  Ability to turn the orientation in any direction
tions (e.g, monitoring and evaluating sensor data, or con- This allows the robots to get from a location (x1 ; y1 ) and
trolling an actuator). Usability requirements dictate that an orientation 1 to a new location (x2 ; y2 ) and orienta-
the benchmark application components be dynamically tion 2 by following a trajectory while turning continu-
controllable, relocatable and replicable. Traditional com- ously. Besides the base, our robot is equipped with a ball
puting benchmarks [15], [7], [19], [16] are inadequate kicking mechanism and a dribble bar.
for characterizing dynamic distributed real-time systems. The controls part of the robot can be classified into
Primarily, they were not designed to exhibit behavior two main components - the microprocessor and the mo-
characteristic of control systems. Also, they focus on tor related electronics. Figure 2 shows a schematic of
component level benchmarking, and thus they do not our control system. The player robots do not require a

0-7695-0990-8/01/$10.00 (C) 2001 IEEE

Authorized licensed use limited to: UNIVERSITY OF SYDNEY. Downloaded on April 05,2010 at 23:37:28 EDT from IEEE Xplore. Restrictions apply.
 3

for resource management. Dynamic resource allocation

Microprocessor
is at the core of high-performance network operations. It
Inverse
Work
Station
Path Planning Kinematics
Compensation
involves the real-time management of communication re-
+
−
sources, that is, intelligent allocation of frequency chan-
nels, spreading-codes and transmission time-slots to net-
Motor/Load
work users, in order to share the scarce radio bandwidth
Actuator
Dynamics
efficiently among them.Wireless LANs typically oper-
Encoder
ate on a dedicated bandwidth, which for several of the
Motor and Electronics most popular frequency allocations ranges from 26 to
CAMERA
150 MHz. One of the most popular frequency bands is
that centered at 2.4 GHz. The channel capacity can be
Fig. 2. Control System Block diagram effectively utilized if the channel allocation is dynamic
[4]. This is taken care of by agents that dynamically ad-
just the characteristics of the data stream to match the
lot of computational power. Most of the computation is
changing network conditions.
performed on off-board workstation. The off-board com-
puter analyzes the visual information, makes predictions
B. Software Design
about the player movements and then tells each robot
where he is and where he must go. The processing each The architecture proposed by us is applicable to both
robot needs to do is: the simulator league and the small sized robot league.
 Given a final velocity vector, orientation and posi- The existing RoboCup client models do not adequately
tion, compute and execute the minimum time path take into account the real-time constraints faced by the
from the current velocity vector, orientation and po- soccer agents on the field. This leads to an overworked
sition and sub-optimal system. To perform these activities un-
 Given a path, follow the path with a given speed. A der real-time conditions, an architecture is proposed that
path is a series of positions/orientations with conti- puts the entire RoboCup subsystem under the control of
nuity conditions specified at each point (This is to a resource manager. This manager monitors the perfor-
allow slower speed motions along a well-specified mance of various players on the field and balances the
path) processor load based on their capacities. In the pro-
 Keep track of his position, speed and orientation be- posed model, we introduce the concept of a Client man-
tween vision updates ager that resides in a level between the server and the
 Decode/process communications. The off-board clients. This client manager, based on the input from the
computer may send messages to specific robots to resource manager will decide the resources needed by
the effect of arming the kicker, or activating it, de- every client. The client manager communicates with the
pending on how the kicker is implemented server using the libraries native to the server and with the
 Control the speed of the spinning blade robot kicker clients using DeSiDeRaTa’s communication libraries.
In a robot soccer system, the robots need to commu- Our architecture is a layered one. This is similar to a
nicate with each other and with the base station. This subsumption architecture [3], [1]. Figure 2 illustrates this
communication needs to be done frequently to help robot architecture and the various layers. Level 0 has only two
share plans and find their goals. Most of these systems functions - detect/avoid obstacles and gather statistics.
use a base station or controller. Wireless LANs allow These statistics indicate the number of collisions that oc-
the robots to transfer information. There are two ways to curred and the number of exceptions sent back to level
connect the wireless LAN to the hardware on board. In 1. If there are no obstacles, the player continues on his
our case, the processor on board the robots has a PCM- current path. Otherwise, he throws an exception to level
CIA interface for the wireless LAN. In some cases, there 1. Level 1 deals with the factors that affect the action and
is a micro-controller on board. For these kind of systems, performance of the client. This includes object tracking
there are wireless serial modems that connect to the se- and exception handling from level 0. Level 2 pertains
rial port on board. This module provides a lot of avenues to the individual planning adopted by the clients. Some

0-7695-0990-8/01/$10.00 (C) 2001 IEEE

Authorized licensed use limited to: UNIVERSITY OF SYDNEY. Downloaded on April 05,2010 at 23:37:28 EDT from IEEE Xplore. Restrictions apply.
 4

main disadvantages with edge detection techniques are:

SOCCER SERVER
 They don’t take advantage of the color information
Player Actions Sense Information
for identifying objects
CLIENT MANAGER
(Resource Manager)
Dash  They are sensitive to noise

Request for Command

Suggested
Command Turn Thresholding [9] is readily applied to color images, and
Sense Info
Level 2
Level 1
it does take advantage of the color information. Objects
Level −1
Command Library
Level 3 Exceptions
in an image are distinguished by their color. Pixels with
Level 0
Obstacle
similar colors are grouped together by comparing each
Avoidance
pixel in the image to a set of threshold values. Thresh-
olding suits the robot soccer environment well, because
different types of objects have different colors. Also, this
Fig. 3. Layered Architecture method is very efficient, since comparing an image to a
set of color ranges would be faster than edge detection.
The task of the vision system is to supply the strategy
of the actions we identified under this category were the
actual strategy planning (offense, defense and goalie), module with information about the position of the objects
player communication, utility of reasoning and predic- it can see relative to a given robot. After thresholding, the
tion. In level 3, we perform the action evaluation of the point on each group of pixels closest to the center of the
clients. This includes modeling the opponent, sharing the image is identified. This gives us the distance between
knowledge learned and computing the resources needed various points in an image. Since the camera is cali-
to reason. All the levels following level 3 would follow brated, we know the correlation between the distances in
the same communication path. The main advantage of the image and the actual distances between objects on the
this scheme is that the clients may be distributed across field. This would enable a robot to know the whereabouts
different machines. The resource manager would still be of every object on the field. However, if some players
are blocking the field landmarks and the ball, estimating
managing both the CPU and the network loads.
the relative position and orientation becomes difficult. In
The objective of a vision system is to capture frames
such cases, the players of a team would need to cooper-
from an overhead camera and extract information from
ate and belief networks [10] would be used to resolve the
the picture for our strategy module. That involves rec-
uncertainty.
ognizing objects in the image (like the ball, or another
In the case of other sensing strategies, the agent should
robot) and determining where they are in the world co-
find the ball and identify the target. Besides vision, typi-
ordinate system. Robot soccer is a difficult environment
cal sensors used in mobile robot research are range find-
for image processing. Primarily, it has to be done in real-
ers, sonars, and bumper sensors. However, it is difficult
time so that the game has not changed significantly when
for them to discriminate the ball and the target unless
the robot is ready to react to the input from the vision sys-
special equipment, such as a transmitter, is set inside
tem. To meet this demand, processing one frame should
take less than (1=30) of a second. Another issue is the
the ball or target, or a global positioning system is ad-
ditionally used to inform the positions of all agents. The
size of the camera. Besides the overhead camera, the
simplest case is a global positioning system, without on-
goalie robot is equipped with an on board camera. Our
board sensing.
on board camera has to be small and light to fit on the
robot. Small and light cameras tend to have low reso-
IV. C ONCLUSIONS
lution and unstable color reproduction. There are two
major techniques that allow us to identify the location Resource management is vital to dynamic systems op-
of objects from an image. They are edge detection and erating in real-time. This paper describes a testbed being
thresholding. Edge detection [8] recognizes edges in the developed for evaluation of resource management mid-
picture, and takes into consideration a priori knowledge dleware. We have described the various modules that
about the shapes of the objects on the field. Edge detec- constitute the RoboCup system and discussed teh ap-
tion is primarily a technique intended for greyscale im- plicability of the resource manager to a few of them.
ages, but can be adapted to color images too. The two Our system will allow assessment of how effectively re-

0-7695-0990-8/01/$10.00 (C) 2001 IEEE

Authorized licensed use limited to: UNIVERSITY OF SYDNEY. Downloaded on April 05,2010 at 23:37:28 EDT from IEEE Xplore. Restrictions apply.
 5

source managers balance the load between the different [16] R.P.Weicker. An overview of common benchmarks. In IEEE
machines running the clients so as to ensure real-time Computer, volume 23, pages 65–75, December 1990.
[17] Ryan Detter, Lonnie Welch, Barbara Pfarr Brett Tjaden and
constraints and achieve the most efficient allocation of
Eui-Nam Huh. Adaptive management of computing and net-
computational resources. The agents being developed by work resources for spacecraft systems. In The 3rd Annual Mili-
us are fully cognizant of the resources available to them tary and Aerospace Applications of Programmable Devices and
and the constraints of operation. These agents will effi- Technologies Conference, September 2000.
[18] Lonnie Welch and Behrooz A.Shirazi. A dynamic real-time
ciently controlled by a meta-agent which manages their
benchmark for assessment of QoS and resource management
time and resources. technology. In The IEEE Real-Time Technology and Applica-
tions Symposium, pages 36–45, June 1999.
R EFERENCES [19] W.J.Price. A benchmark tutorial. In IEEE Micro, pages 28–43,
October 1989.
[1] Eyal Amir and Pedrito Maynard-Reid II. Logic based subsump-
tion architecture. In Sixteenth International Joint Conference on
Artificial Intelligence, 1999.
[2] T.P. Baker. Stack-based scheduling of real-time processes. Jour-
nal of Real-time Systems, 1991.
[3] Rodney A. Brooks. A robust layered control system for a mo-
bile robot. Technical report, MIT, 1985.
[4] D. C. Cox and D. O. Reudink. A comparison of some chan-
nel assignment strategies in large-scale mobile communication
systems. In IEEE Transactions on Communications, volume 20,
pages 190–195, April 1972.
[5] J.S. Breese E.J. Horvitz and M. Henrion. Decision theory in
expert systems and artificial intelligence. International Journal
of Approximate Reasoning, pages 247–302, 1998.
[6] Yasuo Kuniyoshi Hiroaki Kitano, Minoru Asada and Itsuki
Noda. RoboCup: A Challenge Problem for AI and Robotics,
volume I of RoboCup 97: Robot Soccer World Cup, chapter 1,
pages 1–20. Springer-Verlag, 1998.
[7] H.J.Curnow and B.A.Wichmann. A synthetic benchmark. In
Computer Journal, volume 19, pages 43–49, January 1976.
[8] J.Canny. A computational approach to edge detection. IEEE
Transactions on Pattern Analysis and Machine Intelligence,
1986.
[9] William K.Pratt. Digital Image Processing. Wiley-Interscience,
1991.
[10] Hongjun Li. An introduction to belief networks. Technical Re-
port TR 99-58, University of Maryland at College Park, 1999.
[11] C.L. Liu and J.W. Layland. Scheduling algorithms for multi-
programming in a hard-real-time environment. Journal of the
ACM, 1973.
[12] Lonnie Welch, Binoy Ravindran, Behrooz Shirazi and Carl
Bruggeman. Desiderata: Qos management technology for
dynamic,scalable,dependable,real-time systems. In The 15th
Symposium on Distributed Computer Control Systems, pages
7–12, September 1998.
[13] Lonnie Welch, Binoy Ravindran, Behrooz Shirazi and
Carl Bruggeman. Specifications and modeling of
dynamic,distributed,real-time systems. In The IEEE Real-Time
Systems Symposium, IEEE Computer Society Press, pages
72–81, December 1998.
[14] Masayuki Ohta. Learning Cooperative Behaviors in RoboCup
Agents, volume I of RoboCup 97: Robot Soccer World Cup,
chapter 8, pages 412–420. Springer-Verlag, 1998.
[15] R.P.Weicker. Dhrystone: A synthetic systems programming
benchmark. In Communications of the ACM, volume 27, pages
1013–1030, 1984.

0-7695-0990-8/01/$10.00 (C) 2001 IEEE

Authorized licensed use limited to: UNIVERSITY OF SYDNEY. Downloaded on April 05,2010 at 23:37:28 EDT from IEEE Xplore. Restrictions apply.