Fullscreen HTML5Med Res (??MB)HTML4

Roboc Machines playing soccer? Better believe it! Since 1997, teams from across the world have been competing to build the perfect soccer-playing robots, in hope of one day developing the technology to such a level that it may compete against, and beat, the human world cup champions. 8  Silicon Chip siliconchip.com.au cup by David Perry T he challenge for roboticists is developing machines that may think and act autonomously, able to analyse sensory data and act in a meaningful way, as a human would. It was once thought that getting robots to act intelligently in the real world would be a fairly trivial step beyond computer simulations of such an environment. Several decades later, we now appreciate the complexity of the tasks that nature manages with such apparent ease. Despite millions of dollars, countless postgraduate students, academics and corporations, robots can only begin to operate intelligently within a tightly controlled environment. Any attempt to generalise a vision or learning algorithm into something mildly human-like is met with little success and a great amount of frustration. There are several factors which are making this technology evolve slower than we’d like. One is the inter-disciplinary nature of robotics and A-I (henceforth just called intelligent robotics). Robotics researchers may have a background in any number of engineering and scientific fields: computer science, cognitive sciences, mechanical, electrical or software engineering. Everything breaks – and interacts! I like to think that robotics involves Murphy’s Law cubed. The mechanics break, the electronics break and the software breaks! There are all kinds of interactions within and between these systems, challenges spanning multiple fields of engineering and sometimes aspiring to emulate the natural world. Computer-based systems can evolve because there is some standardisation amongst hardware and software. Everyone knows about C++ and has access to (and uses) PCs. But complete off-the-shelf solutions are not common in robotics. Researchers often begin by building their own mechanical platform, selecting and designing wheels, legs, motors, the chassis... Then they build motor controllers, select batteries, design controllers or work out how to use a PC or notebook computer in their design. Only a handful of companies build wheeled robots (or humanoid robots) for sale to researchers. And then they tend to be expensive and not optimal for particular research interests. In building robots, there are just too many design considerations, technology that needs to be implemented but which is not yet mature. Everybody goes off on their own research tangent and, as such, tend to develop everything from scratch, by themselves. Soccer is the equaliser One aim of the RoboCup Federation is to overcome these difficulties by providing a standard, scalable problem. That standard is the game of soccer, with a number of “leagues”. The challenge begins with playing soccer in simulation and in well lit environments with well-defined, coloured markings. The rules slowly evolve so that as the environment rules become less stringent, the robots become more robust, adaptable and, hopefully, more intelligent. Every year, the rules of each league are reviewed with aim of encouraging innovation. Field walls have been removed in some leagues and eventually colouring and lighting requirements will be relaxed. The idea is, once the ultimate challenge is fulfilled, robots will have matured to such a state that they will be feasible in everyday life. Of course, there is some way to go. The most developed robotic league in RoboCup is called the F-180 or small-size league. The ‘180’ denotes that robots may have a footprint no larger than 180cm2 – although the siliconchip.com.au May 2004  9 The F-180 league is relatively mature, games are extremely fast and teams implement complex behaviours including passing and blocking. Middle-size league Depending on the league, control can be as simple as one notebook computer – or a computer for every robot. Most work on wireless LANs. actual sizing requirements are now more complex. The game is played on a greencarpeted field about the size of a pingpong table. One goal is yellow, the other blue and the field is surrounded by a small ramped barrier which will keep a ball in play, provided it hasn’t been struck too hard. This league allows teams to use an overhead camera, connected to as much computing power as a team cares to bring with them. The computer can see the ball (an orange golf ball), the goals and every team member at all times. Local on-board vision is also allowed, although teams that use this exclusively tend to be not all that competitive. The robots themselves may have a few sensors on board (for example infrared beam sensors to detect when the robot is in possession of the ball) and some low-level control electronics. The robots are commanded by RF from the host computer which outputs the behaviour for all of the robots. The robots are in teams of five, including a goal keeper. The drive system typically entails either a differential drive (like a tank), or (more recently) omnidrive, which allows the robot to move in any direction, from any orientation. So where a robot with differential drive has to turn to the direction it wishes to travel, omnidrive can move in that direction instantaneously. It achieves this using three or four omnidirectional wheels. As well as turning like a normal wheel, these 10  Silicon Chip have rollers that allow the wheel to move in the direction of its wheel axle. By varying the speeds of the wheels, the resultant drive and slip of the rollers allows the robot to move in any direction. These robots also often include ball manipulation devices, including a kicker (usually a spring-loaded or solenoid mechanism) and ball dribbler (a rotating rubber coated bar on the front of the robot that induces a back spin on the ball, keeping it close to the robot even when travelling backwards). It would be a reasonable assumption that a computer controlled F-180 team could defeat a team of humans playing with joysticks (at least until the humans got in a great deal of practice). The F-2000, or middle-size league is less mature. Like the F-180 league, the ‘2000’ denotes that the robots may have footprints no larger than 2000cm2. Their basic physical construction is similar to that of the F180 league, only scaled up. Bigger batteries, motors and beefed-up control electronics are required to deal with this additional load. This increases cost dramatically. Their game is played on a larger field (8 x 12 metres) with a similar colouring scheme to the F-180 league. There is now only a white line at the field’s edge, so the ball can be kicked out of play easily. At each corner of the field is a coloured post. What makes this league far more challenging is the requirement for all sensors and processing to be onboard (an off-field ‘coach’ computer is permitted). The robots must now play soccer, not always having a view of the ball, team-mates, nor the goals. With the F-180 league, you could get by with one cheap desktop computer but F-2000 needs one for each robot, powered by batteries – either modified desktops, single board computers, or laptop computers. Vision consists of one or more cameras on each robot. Some use cameras aimed at a convex mirror This close-up give a good idea of the complexity involved in a soccer robot. siliconchip.com.au mounted above the robot, providing a panoramic view of the field. While this has the advantage of giving the robot the ability to see in all directions, the detail that can be viewed in a particular direction, especially at a distance, is reduced. The standard of play in this league is much lower. Given the size of the field, robots seem to move more slowly. They tend to get caught on things, run out of bounds and can’t manipulate the ball with the same agility as smaller robots. The robots lack localisation (working out where they are) to the same accuracy and thus are limited in the complexity of strategies able to be taken. Successful robots in this league tend not to be the fastest or most clever but well tuned to guarding their own goal and able manoeuvre a ball around obstacles towards the opponent’s goals. Humanoid league The challenge takes a final step up in the humanoid league. These will be the robots that will hopefully face humans in a full-scale match in 46 years (2050 is the RoboCup organisation’s target date). At the moment, however, they’re quite primitive. Robots can be in a number of size classes, some over 2m tall. The tasks they must complete include subsets of skills needed for an actual game (walking, penalty kicks, standing on one leg) and a one-on-one competition. These robots can be enormously expensive to build. While wheeled robots need only a few actuators, humanoids can have dozens. The power requirements are considerable, as are the computational requirements (humanoid robots require very high speed control systems that can deal with balancing the robot in real time). The development of humanoid robots is a story in itself. Honda spent decades and hundreds of millions of dollars developing their humanoid robots, ASIMO being the latest. These intensely engineered robots can make the whole task seem easy. They’re often described as looking and walking like men in space suits. Some robots in this league do have this level of agility and apply it in playing soccer. Game play at its best, however, is still very slow and delicate. The robot’s foremost concern is siliconchip.com.au This shows the long and the short of it: some of the humanoid-class robots stand 2m tall and cost a fortune to build . . . staying upright so any attempt to kick or block a soccer ball comes second to maintaining balance. Other leagues There are four other leagues in the competition which complement the F-180, F-2000 and humanoid competitions. The first of which is a simulation league, where competitors create teams entirely in software. Unhindered by hardware, they can focus on creating intelligent agents which are able to perceive environmental data provided by the ‘soccer server’ and formulate the necessary actions to be taken. Software teams are played off against each other in the “virtual arena” of the soccer server and the game is shown on large screens throughout the competition. The rescue league diverts completely from the game of soccer, aiming to provide an immediate practical application. There are two components to this league. The first is a simulation league where the behaviour of emergency responses (for example fire fighters) is simulated on a large scale. It is hoped this research will provide data to assist emergency services in best distributing resources in disaster scenarios. The second consists of real rescue robots that may be sent into a disaster zone that is either inaccessible or too dangerous for human rescuers to reach. The robots are varied, some having on-board sensors and intelligence. Others are tele-operated, either by radio or a cable dragged behind. In competition robots are sent through a . . . on the other hand, there are little fellas in the RoboCup competition too. May 2004  11 They’re cute to watch: the four-legged league is based on Sony’s “AIBO” robot dogs and creates a level hardware playing field. course of unknown configuration, and must avoid obstacles to find and map the location of potential survivors. The ‘survivors’ are plastic mannequins which may have actuated limbs that shake, or heating pads to give the impression of a warm, alive body. The robots detect the survivors through vision, thermal detection or motion sensing. Similar robots were deployed at the World Trade Centre to search through the debris for survivors, unfortunately without success. The scoring is dependent on the quality of data acquired and the amount of human assistance. In the future, the league will be expanded so that robots can perform additional rescue functions, such as administering first aid or reinforcing unstable structures. Perhaps the cutest robots belong to the four-legged league. Sony AIBO robotic dogs are programmed to play soccer, on a field similar to that of the F-180 league. Sony is a major sponsor of RoboCup so the league is good publicity. But the fixed hardware also forces teams to deal entirely in software. In some ways it creates a fairer competition, as teams can’t simply buy more powerful motors or computers to defeat the competition. Every effort is made to extract maximum performance from the fourlegged dogs. Extensive fine tuning of the walking gait ensures maximum speed and the robots often walk with their heads low to the ground and legs spread out to block or intercept the ball. The Sony AIBO ERS-210A robot, 12  Silicon Chip one of the newer versions, has 20 degrees of freedom (DOF) throughout its four legs, head, ears and tail. The computing power on-board is roughly equivalent to a small PC or PDA, with its 64 bit RISC processor running at 384MHz. The robots typically make use of wireless networking, as with other leagues, to communicate and also take game start/stop instructions. The legged league may be cute and fun to watch but the game-play is still of relatively low quality. Despite the valiant efforts of their programmers, the robot’s hardware is limited. The single on-board CMOS camera has limited resolution (176x144), the motors have limited torque and speed, and the computer has limited power and memory. The situation will improve, with the latest version of AIBO, the ERS-7. This robot should make its debut at RoboCup 2004. It may well be that the league advances along with Sony’s periodical release of newer, more advanced robots. As with the main competition there are several leagues, including soccer, rescue and dance, each with set age groups. The soccer league is either 1-on-1 or 2-on-2. Several changes are made from the grown-up version to simplify the technology required to compete. For example, the field has a monochrome gradient, rather than green carpet. By reading the field with a light sensor, a robot can judge its position along the length of the field. The goals are black at one end, and white at the other, also allowing them to be found with a simple light sensor rather than an expensive camera. There are 140mm high walls around the edge of the field. Finally, the ball itself is a clear plastic shell with several infrared LEDs inside. Thus the ball can be found by scanning for the unmodulated infrared light emitted, using a filtered light sensor like a photodiode. The junior rescue league is essentially a scaled down version of the parent. The course is smaller and has less obstacles. There is a line that a robot may follow that will lead it through all the various sections of the course, without the need for complex navigation. The survivors are denoted by either a green or silver patch on the floor of the course. These colours can be differentiated with a light sensor, aimed at the floor. In the dance league, aimed at younger students, robots are constructed and Who’s playing – and paying? The majority of participants in RoboCup are universities, with a small number of companies also putting forth teams. The costs of parts, travel and registration don’t leave much room for the hobbyist. However there is a concession in the form of RoboCup Junior, a branch of the competition for primary and secondary school students. Where RoboCup focuses on fostering research, RoboCup Junior is about encouraging education in science and technology. In every RoboCup competition arena you’ll find a “pit” area set up for last minute tweaks and of course repairs. siliconchip.com.au programmed to dance to a song for up to two minutes. Some have clever motorised limbs and timing, others are simply dressed up. Robots were at first solely constructed using Lego Mindstorms. These Lego systems have sensors and controllers well suited to the junior soccer, rescue and dance leagues. Drive systems are also quite easy to construct. Advanced teams now use more customised hardware, including microcontroller boards and ultrasonic sensors. Melbourne Uni team I was involved with the University of Melbourne mid-size team, within the Department of Electrical and Electronic Engineering. We competed at RoboCup2003 in Padua, Italy. It was the first year we had competed. This, combined with our extreme under-preparedness, proved extremely challenging. Due to delays in manufacture, much of the hardware was untested, up to the point that we were assembling things for the first time, just days prior to the competition (even on the plane to Italy!). The MU-Wallabies, as we were known, had several design criteria in mind. One was the need for small, highly manoeuvrable robots. Also, an omnidirectional kicking device was designed. It consisted of a ‘leg’ which was able to be rotated about the robot’s circular body. Unfortunately, a number of problems arose. In order to fit in a kicking device that protruded from the robot’s body, the body had to be quite small. This required that the wheels (we used differential drive) be close together and that the laptop computer used to control the robot be mounted vertically. Ooops! The first question we were asked about our robots, after being asked whether they could participate in “Robot Wars” (instead of Robocup!), is whether they will tip over. We were hoping software algorithms would prevent this from happening, although sometimes collisions or malfunctions would cause a robot to turn too tightly and fall over in competition – much to the opposing teams delight. Also, the aim of having fast robots has yet to be fulfilled. The cheap motors and gearing used failed to perform to expectation and we’ve been forced siliconchip.com.au to move to much more expensive precision Swiss-made motors with planetary gearboxes. Our entire drive system, despite requiring custom CNC machined parts, only cost around $1500 per robot. The new motors alone will cost $1200 per robot. Additional problems included faulty wiring and overheated battery packs. All of these things were quite preventable but we simply lacked the time to catch all the bugs. The laptops run a Slackware Linux distribution, and the C++ control code is compiled with GNU (open source) tools. Inexpensive Logitech webcams are used for vision, through the USB port. They run on the Video4Linux drivers that are compatible with the Philips webcam chipset. Images of 320 x 240 pixels are captured and processed at a speed of about 10 frames per second. This maximum speed is a limitation of the webcam, and not the software or laptop. The software classifies regions in the image based on colour and makes a determination as to what the corresponding object is. It needs to be calibrated, as colours can vary depending on lighting conditions. The vision software provides polar coordinates (angle and distance) for each object in view on the field. The next step in processing is either acting directly on this data, or using it for localisation. This process determines where a robot is in relation to the entire field. For instance, it may output a cartesian x-y co-ordinate, where the origin is in the centre circle. Through the 802.11b wireless network, robots can collate this information and build a detailed model of where everything is on the field. Artificial intelligence software takes this information about the environment and works out what actions to take. In 2003, this software was extremely simple. That said, it did work surprisingly well and our main source of problems was hardware breakdown. Australia is a strong, competitive participant in the RoboCup competition. Other institutions competing include the University of Queensland, University of NSW, University of Newcastle and University of Technology, Sydney. In 2003, researchers from Australia took out both the Engineering Challenge Awards. rUNSWift from the University of New South Wales took first place in the Sony legged league, with the Nubots from the University of Newcastle coming in third. One of the most exciting matches saw the RoboRoos of the University of Queensland facing BigRed from Cornell University in the US, in the F-180 final. Cornell has enormous financial and human backing, including Microsoft and NASA, and they have dozens of students and staff assisting. Despite this, they were defeated in an earlier round-robin match by the University of Queensland. Unfortunately, Cornell managed a 1-0 defeat of the RoboRoos in the final, with the match having gone into overtime. RoboCup 2004 is to be held in Lisbon, Portugal, from June 27 to July 5. If you would perhaps like to get involved, or just watch these robots in action without the expense of overseas travel, there may be a local alternative. RoboCup Junior has state and national competitions every year. For the more ambitious, the University of Melbourne mid-size RoboCup team is aiming to set up a wider robotics competition later this year that will SC be open to all. Weblinks: www.robocup.org/ – RoboCup Official website www.robocup2004.pt/ – RoboCup2004 ww.robocupjunior.org.au/ – RoboCup Junior Australia robocup.ee.mu.oz.au/ – University of Melbourne mid-size team May 2004  13