Real-time Control of Soccer-playing Robots

Using Off-board Vision: the Dynamite Testbed

Michael K. Sahota, Alan K. Mackworth, Stewart J. Kingdon and Rod A. Barman

Laboratory for Computational Intelligence

Department of Computer Science
University of British Columbia
Vancouver, B.C., Canada, V6T 1 2 4
dynamo @cs.ubc.ca

ABSTRACT it is a unique and successful approach to the trade-

offs involved in mobile robot design although a re-
The Dynamite testbed has been constructed for ex-
lated scheme was independently proposed in [5].
periments with autonomous mobile robots. It con-
sists of a fleet of radio-controlled vehicles that re- 2. ROBOTS
ceive commands from a remote computer. All the
robots share an off-board vision system that tracks The mobile robot bases are commercially available
the position and orientation of each robot at 6OHz. radio-controlled vehicles. We have six 1/24 scale
A novel aspect of this system is that computer vision racing-cars, each 22cm long, 8cm wide, and 4cm
is used not only for sensing the environment, but high excluding the antenna. Although we plan to
also for sensing the state of the robot itself. The for- use all the cars, the most that have been used,
mer information is used for planning and the latter so far, at any time is two. The testbed (244cm
for control. The testbed has been used successfully by 122cm in size) with two cars and a ball is
for experiments with robot soccer games to compare shown in Figure 1. The cars have each been fitted
the effectiveness of altemate control architectures. with two circular colour markers allowing the vision
Testbed use is facilitated through a graphical simula- system to identify their position and orientation.
tor that allows control programs to be tested off-line The ball is the small object between the cars. The
at a workstation. The testbed has been developed to robots are neither as flexible nor as competent as
the point where we now have a portable version that human soccer players, so we have modified the
has been demonstrated at a number of conferences. environment in two ways. First, there is a wall
around the soccer field which prevents the ball, and
1. INTRODUCTION the players, from going out of bounds. Second,
there are barriers to prevent the ball from getting
The Dynamite testbed provides a practical platform
trapped in the corners. Since these are Canadian
for testing theories in the soccer domain using mul-
robots, it is not unreasonable for the soccer field to
tiple mobile robots. The testbed consists of a fleet
be shaped like an ice hockey rink.
of radio-controlled vehicles with perceptual, plan-
ning and control systems. The robots perceive their
world through a shared off-board perceptual sys-
tem. In an integrated off-board environment with
dataflow and MIMD computers, vision programs
monitor the position and orientation of each robot
and the ball while planning and control programs
generate and transmit motor commands. This ap-
proach allows umbilical-free behaviour for light- I-

weight fully autonomous robots. As far as we know,

Figure 1 . Robot Players on the Soccer Field

0-7803-2559-1195 $4.00 0 1995 IEEE 3690

 ! q Z b r H
Lookup
Table
, f-i ,
Encoder J
Runlength

T805 Transputer
(Microprocessor)

Llnksto other ?, ? ?, ?,
Figure 2. Hardware Configuration Transputers
+ i i i
I

3. SYSTEM ARCHITECTURE Figure 3. Architecture of the Colour Tracking System

The original architecture of the testbed [l] is

shown in Figure 2. There is a single colour camera MIMD computer. Additional vision processing is
mounted in a fixed position above the soccer field. performed on the transputers to find the position, in
The output of the vision system is common to all screen coordinates, of the centroid of each coloured
controllers. The different visual tracking hardware blob and to transform these positions from screen to
systems and algorithms are described in the next world coordinates. The vision subsystem is called
section. The reasoning and control components of a the Vision Engine [7]. The Vision Engine produces
vehicle can be implemented on any number of trans- the absolute position of all the objects on the soccer
puters out of the available pool. Currently, each field; the orientation of each car is also reported.
vehicle is controlled by a two-process user program This is done at 60Hz with an accuracy in position
running on a single transputer node. An arbitrary of a few millimeters.
number of nodes, labeled 1 to n in Figure 2, can The second vision system accomplishes the
be used in parallel to control independent vehicles. colour classification through the use of an analog
The movement of all vehicles is controlled through circuit and performs the remaining computation on
radio transmitters attached to a single shared trans- transputers as before. The analog circuit thresh-
puter node. Commands are transmitted to the vehi- olds the video signal to provide a coarse division of
cles at a rate of 60Hz. colour space so that only red and green targets can
A feature of the Dynamite testbed is that it is be tracked [lo]. This system is portable since all
based on the “remote brain” approach to robotics of the required computer hardware fits in a chassis
161. The testbed avoids the technical complexity that is the size of desktop computer.
of configuring and updating on-board hardware and The third system uses a digital colour video
makes fundamental problems in robotics and artifi- processing module, called the Colour Tracking Sys-
cial intelligence more accessible. tem (CTS), which was developed to perform the
colour classification. The main functional groups
4. VISION of the CTS hardware are shown in Figure 3.
The CTS is able to grab RGB video input in a
The basis of the vision system is the tracking of variety of modes. Each pixel is classified, using the
coloured discs that are placed on the robots. Even Iookup table, into one of a number of classes based
for this simple visual task, specialized hardware on its colour attributes. The runlength encoder com-
is required. Three vision systems have been con- presses the description of the classified pixels so
structed that use the same algorithms but run on that they can be processed by the microprocessor
different hardware. in real-time. The hardware that comprises the run-
The original system uses DataCube video pro- length encoder is reprogrammable and allows the
cessing hardware. The DataCube is a dataflow functionality of the CTS to be altered without hard-
processor which has been programmed to classify ware modifications. The INMOS T805 transputer
image pixels into different colour classes at 60Hz is a general purpose microprocessor that segments
(interlace video rate). This information is trans- the compressed classified pixels into connected re-
mitted to a network of transputers which form a gions and communicates this information to other

3691
 partitioned into a deliberator and an executor; the
distinction is primarily based on the different time
scales of interaction. Informally, the deliberator de-
cides what to do and how to do it, while the ex-
ecutor interacts with the environment in real-time.
These components run asynchronously to allow the
executor to interact continuously with the world and
the deliberator to perform time consuming computa-
tions. This partition is inspired by recent architec-
tures that attempt to integrate planners with more
reactive components [3; 41 as well as architectures
Figure 4. The Dynamite Simulator with Two Cars and a Ball
such as subsumption [2].
Several controllers based on reactive deliber-
transputers in the system. The CTS conforms to the ation have been implemented to allow robots to
Transputer Module (TRAM) daughter card standard compete in complete one-on-one games of soccer.
which allows it to be used in a variety of comput- A series of experiments called the LCI (Labora-
ing environments. tory for Computational Intelligence) Cup has been
The speed of the system depends on the num- conducted to compare alternate controller designs.
ber and size of coloured objects present in the cam- The soccer toumament provided evidence that the
era’s field of view since the amount of computation features proposed for our robot architecture Reac-
performed by the the transputer increases with the tive Deliberation are useful. There are three main
number of classified pixels to be segmented. For results. First, intelligent action selection cannot
the Dynamite Testbed, the CTS is easily capable of proceed independent of planning in non-trivial do-
60Hz operation with a latency of several millisec- mains. Second, goal-oriented behaviours are use-
onds following the reception of the video field. ful abstractions that allow sharing of scarce com-
putational resources and effective goal-arbitration
5. SIMULATION
through inter-behaviour bidding. Third, experimen-
A physics-based graphics simulator for the Dyna- tal evidence has been provided that the goals and
mite world is available for testing and developing actions of a robot need to be evaluated at a rate
reasoning and control programs. It has proven to commensurate with changes in the environment.
be very useful in the development of control pro-
grams. The simulator supports an arbitrary number 8. CONTROL
or cars and balls, as well as a simple contact model The deliberator is responsible for selecting an ap-
for collisions. A visual display of the simulator is propriate action or goal in the current situation.
shown in Figure 4. This includes the problem of non-holonomic path
planning since the most appropriate action may de-
6. PLAYING SOCCER
pend on how long it takes to move from the cur-
Soccer has been proposed as a task for the develop- rent position to another location. The executor for
ment and unification of divergent theories in Artifi- the soccer-playing robot consists of five action low-
cial Intelligence [9]. Soccer captures many impor- level schemas: follow path, servo at ball, defend,
tant aspects of the real world such as competition stop, and idle. Only one action schema is enabled at
and cooperation with other agents, a dynamic envi- a time and the enabled schema sets the robot’s con-
ronment and real-time interaction. trol outputs (throttle and steering angle) and sends
messages to the active behaviour in the deliberator
7, REACTIVE DELIBERATION when it is having problems.
A control architecture called Reactive Deliberation The current functionality of the controller
has been proposed for the task of soccer playing [SI. includes various simple offensive and defensive
Under reactive deliberation, the robot controller is strategies, motion planning, ball shooting and play-

3692
ing goal. The robots can drive under accurate multiple mobile robots. In Proceedings of Intelligent
control at speeds up to I d s , while simultaneously Vehicles Symposium (1993), pp. 261-266.
considering alternate actions. We have produced a [2] Brooks, R. A. A robust layered control system
10 minute video that documents these features. for a mobile robot. IEEE Journal of Robotics and
Automation RA-2 (1986), 14-23.
9. CONCLUSIONS
[3] Firby, R. J. Building symbolic primitives with
The features and capabilities of the Dynamite continuous control routines. In First International
testbed have been described in detail. The testbed Conference on Artijicial Intelligence Planning Sys-
has been used extensively for experiments in the tems (1992), pp. 62-69.
soccer domain. The experiments allowed the com- [4] Gat, E. Integrating planning and reacting
parison of alternative controller designs that has in a heterogeneous asynchronous architecture for
resulted in the Reactive Deliberation architecture. controlling real-world mobile robots. In AAAZ-92
A more practical result is that we have demon- (1992), pp. 809-815.
strated the viability of the remote brain approach
[5] Hallam, J. Playing with toy cars: An experiment
for this class of experiments; the robots we have
in real-time control. Tech. Rep. DAI Research Paper
developed can drive under control at speeds of 1
No. 527, Edinburgh University, 1991.
d s . Off-board computer vision, monitoring both
the robot plant and its environment, can be used [6] Inaba, M., Kamada, T., and Inoue, H. Rope
for real-time robot control. handling by mobile hand-eye robots. In Proc. ICAR-
93 (1993).
9. ACKNOWLEDGEMENTS [7] Little, J., Barman, R., Kingdon, S., and Lu, J.
We are grateful to Dinesh Pai, Jim Little, Heath Computational architectures for responsive vision:
Wilkinson, Steve Monai, Ying Zhang and David the vision engine. In Proceedings of Computer Ar-
Weih for help with this. This work is supported, chitecturesfor Machine Perception (1991), pp. 233-
in part, by the Canadian Institute for Advanced 240. Paris.
Research, the Natural Sciences and Engineering [8] Sahota, M. K. Reactive deliberation: An archi-
Research Council of Canada and the Institute for tecture for real-time intelligent control in dynamic
Robotics and Intelligent Systems Network of Cen- environments. In Proceedings of AAAI-94 (1994),
tres of Excellence. pp. 1303-1308.
[9] Sahota, M. K., and Mackworth, A. K. Can
10. REFERENCES situated robots play soccer? In Proceedings of
[l] Barman, R., Kingdon, S., Little, J., Mackworth, Canadian AI-94 (1 994), pp. 249-254.
A. K., Pai, D., Sahota, M., Wilkinson, H., and [101 Wright, A. A high-speed low-latency portable
Zhang, Y. Dynamo: real-time experiments with visual sensing system. MIT course project, 1994.

3693