At first the scene is reminiscent of any given village football club on a Sunday morning. Goals are erected, white lines mark out a pitch and a ball is placed in the centre circle. But then the players emerge and slowly make their way towards the pitch, not with the befuddled movements of those still feeling the ill effects of the previous night’s immoderation, but with a less fluid, more mechanical gait - awkward stuttering motions that single them out as non-humans. They are robots. And this is RoboCup.

First contested in 1997, the RoboCup is an annual competition that brings together some of the finest minds in robotics and artificial intelligence, primarily to pit small-sided teams of fully automated robots against each other in games of football. It features competitions for robots of varying shapes and sizes, but the one in which kinetic technology and computer programming most equally convene is the Standard Platform League (SPL). 

“I assume that any person of any other RoboCup league would say that their league makes the most interesting advancements,” says Tim Laue, a researcher at the German Research Centre for Artificial Intelligence in Bremen and one of the leaders of the SPL’s successful B-Human team. "In my opinion, the SPL's fascination derives from its unique combination of well-designed humanoid robots, made by a professional manufacturer, and its focus on software development. This allows all participants to concentrate on bleeding edge algorithms, which let these complex robots play soccer on a decent level, without having to struggle with any construction details." 

In the SPL every team uses the same robot, the 57 cm tall Nao produced by the Paris based robotics company Aldebaran. The form of the Nao resembles that of a small human, with all major body parts in the correct positions, albeit connected by just 21 joints. The teams are forbidden from making mechanical alterations or additions to the robot and it is therefore the programming and problem solving skills of the team members that determine the success, or otherwise, of each team.

One of the foremost challenges is coding perception of space and recognition of key objects such as teammates, the ball and the goals. Images are received via two onboard cameras located on the mouth and forehead of the robot and processed using colour and shape recognition algorithms in order to produce a world view. Robots on the same team then communicate with one another over a Wi-Fi network, sharing this information in much the same way as players on a human team shout instructions such as ‘Man On!’ to aid a teammate’s understanding of their surroundings.

“Most robots first collect and process all their sensory inputs, in order to form their current view of the world,” explains Aris Valtazanos, a research postgraduate student at the University of Edinburgh, whose Team Edinferno compete in the SPL. “This can optionally be augmented with the information they receive from their teammates. However, what is more important at this stage is to associate each belief with a confidence value. For example, if a robot sees the ball, then it also knows with high confidence its position relative to it. By contrast, if it sees a single yellow goalpost and no other landmarks, it cannot know its current position with certainty.”

Having established field position and location of key objects, the robot then enters the decision-making phase. “Some example decisions are: ‘Go to the ball or let a teammate that has a better position go?’ or ‘Kick now or position at a better angle?’,” explains Laue. He indicates that the answers and resulting actions are, on the whole, pre-programmed, rather than learned by the robots. “As the current behavior complexity of robot soccer is still quite low, it is, in most cases, easier and more effective to actually program all behaviors explicitly and not to learn any behaviors.”

While fairly robust code in respect of perception, modelling and decision-making is now common place among the SPL teams, the kinetics of the Nao robots are not yet at the same level. At present they are incapable of running and exhibit poor balance when they come into contact with one another, issues that Aldebaran are aware of the long-term need to resolve. “Agility and autonomy are the two main [development] goals,” a company spokesman commented. “It has to be able to play a game without recharging and have the ability to keep equilibrium while running fast."

The 2012 RoboCup saw the debut of Nao 4.0, Aldebaran’s latest, much improved, iteration of the robot. It more than triples the processing speed, from 500MHz to 1.6GHz, quadruples the RAM and offers new cameras with a wider field of view and increased image quality. “The superior quality of the images means that we will now be able to look at more challenging problems we were previously unable to,” explains Valtazanos. “Such as detecting the full pose (position and orientation) of other robots.”

Such improvements were not readily visible at the 2012 competition, hosted at the World Trade Center in Mexico City. Some of the teams had little time to work with the new robots ahead of the tournament, and the competition itself was beset by a various issues, including poor lightning, an erratic Wi-Fi connection that made communication between robots difficult and a vibrating wooden floor that adversely effected the walking motion of a number of team’s robots.

In such circumstances it is perhaps unsurprisingly that the winners were one of the most experienced teams. UT Austin Villa of the University of Texas at Austin are led by Peter Stone, who has been involved in robot football for just under 20 years. “I was a PhD student working on AI planning, but I'd been a soccer player my whole life,” he explains of his introduction to the field. “It was a way to both extend planning to a multiagent setting and to combine my hobby with my research.”

One of the specialist areas of Stone’s research is robust decision making under uncertainty, an area in which his team excelled at the 2012 RoboCup. “We found that our robots made very few mistakes with regards to decision-making,” he confirms. “For the most part, they positioned themselves where we expected, kicked when and where we expected, etc.” This consistency helped UT Austin Villa defeat B-Human 4-2 in the competition’s final, a result that ended a run of three consecutive championships for the German team.

There were elements of UT Austin Villa’s play that suggested passing between team-mates, a tactic rarely implemented due to the potential inaccuracies involved, could in time become a prevalent strategy. “Whenever one of their players went to the ball, the others went, or at least appeared to go, to strategic positions in anticipation of a pass,” Valtazanos enthuses. “From the little I spoke to the Austin Villa people, they said that they had a combination of static, in which a player always executes a fixed behaviour, and dynamic roles for their team players. I thought that was clever.”

It has taken fifteen years to get to the stage where the technology is sufficient to allow such strategies to start to be implemented, but the teams are still some distance away from reaching the long term goal set at the genesis of the RoboCup organisation.

1997, the year of the first RoboCup, is significant as it saw two monumental breakthroughs in the fields of robotics and artificial intelligence: IBM Deep Blue defeated reigning world chess champion Gary Kasparov, and Sojourner, the first autonomous robotics system, was deployed on the surface of Mars.

The RoboCup organisation was formed following discussions between scientists seeking a similarly lofty goal that could act as a stimulus to accelerate research in multi-agent robotics. The idea of robots playing football was first mooted in 1992 and subsequent feasibility studies yielded promising results. In 1995 the RoboCup was announced to the world at the International Joint Conference on Artificial Intelligence in Montreal, Canada, with the stated long term goal that a team of robots would be able to defeat the current human World Cup holders by the middle of the 21st century.

To witness the stilted movement, poor balance and weak kicking of today’s robots and imagine such a target being reached seems a large stretch, but those in the community believe it is not an impossibility.

“If one were to measure the progress that has been made since the start of the RoboCup project about 15 years ago and extrapolate to predict what the state-of-the-art might be like in 2050, then yes, we would probably have a good enough team to beat the world champions,” says Valtazanos. "The main issue is that we do not yet know how to robustly address the hardest problems in the domain, so it is difficult to say if we will be able to do so then.”

Those problems are mainly kinetic according to Laue: "As soccer is played in a very structured environment, most computer science tasks can probably be solved, in our opinion. This might be different for the needed robot actuators.”

“Today's motors and gears are far from being suitable for the intended task,” he explains. “It is not currently known what type of technology is needed to construct something that has the same characteristics as a human leg. Maybe we need something like artificial muscles. Maybe something completely different. However, in my opinion, there has to be a general interest in, or a huge market for, this kind of technology to trigger the necessary advancements. I do not know if that will happen.”

But far from just being a case of whether or not victory is possible, there are other factors that need to be taken into consideration. “No human soccer player would play against a physically superior robot that kicks stronger and runs faster,” says Laue, an issue that was considered in greater depth by Stone in his 2010 paper entitled: ‘The Essence of Soccer, Can Robots Play Too?’

In it, Stone argues that limitations would have to be placed on the physical dimensions, running speed and kicking power of the competing robots so that they do not exceed those of the humans they are competing against. By implementing such restrictions, Stone states that any victory for the robots would come about because “they had better finesse with the ball, they worked better as a team, had better vision with regards to passing, and had better movement off the ball.” Essentially, because they were better football players rather than superior athletes.

Winning in this manner would be of more benefit to the research objectives of the RoboCup organisation as the resulting artificial intelligence advancements could easily be transposed to other, more useful tasks. In addition to football, the RoboCup already includes competitions for search and rescue robots and those programmed for domestic tasks, while Laue explains that certain aspects of his team’s programming, such as navigation and obstacle avoidance, have been used to design an autonomous wheelchair for severely disabled people.

Ultimately it is the quantity of these kind of developments, rather than victory over illustrious human opposition, that will determine the success or failure of the RoboCup project. “Most people only talk about the 2050 goal when they are asked, for example by journalists,” Laue tells me. “In my opinion, it is of no major relevance to our current work.”

Nick is an itinerant freelance football writer for hire, specialising in South America, Mexico and Southern Spain. You can follow him on Twitter @chewingthecoca