Researchers Want Robots to Play in the World Cup by 2050
But first they need to learn not to fall over.
Bold Hearts in 2014. Image: Daniel Polani
Robots are awesome, robots are terrifying. They can jump over hurdles, thrash us at Chess, and withstand radiation, venturing where no human will. But why do they still have so many issues playing a four-aside football game?
"The essential problem is that the robots don't understand what they're doing. They're hardcoded to move from step to step, from action to action, from behaviour to behaviour," Daniel Polani, a UK robot football manager and professor of artificial intelligence at the University of Hertfordshire, told me.
"It's a bit like driving through the fog where you can only see two yards ahead. This is what happens to the robot—they effectively have no big picture of what they're doing."
RoboCup 2015, which is currently holding its 19th tournament in Hefei in Eastern China, involves teams of four robots from all over the world playing 20-minute football matches. The goal of the annual championship is to create robots that are agile, lithe, and robust enough to compete against pro human footballers by the 2050 World Cup. The researchers involved also hope to apply their findings to produce better rescue, disaster, and domestic robots.
Itsuki Noda, the president of the RoboCup Federation and a principal research manager of the National Institute of Advanced Industrial Science and Technology in Japan, told me that it probably wouldn't be hard to eventually engineer an agile robot. The main challenge would be making sure the robot could collaborate with and understand humans.
"Our ultimate goal is to develop a soccer robot that can win against a human champion. It won't be so hard for robots to beat humans [at football]. We will be able to develop a very powerful kicking machine, and super-goalie robot that can move faster than a ball," Itsuki said.
"But a robot should never injure a human, and furthermore, robots need to collaborate with humans, understand human intentions and show their intentions in ways that humans can easily understand. We would like to develop a robot system that can collaborate with humans."
Polani heads up Bold Hearts, a team of four 30-40cm tall robots in the kids-sized humanoid league of the RoboCup. The team were the runners up in last year's finals in Brazil. Polani's teams robots are pre-programmed with around 50,000 lines of code, and as all the robots have to act autonomously on the field, the pre-match preparations are intense. Polani likens the process to rigorous Formula One conditions, where pit stop changes are choreographed down to a tee.
"You can't be sloppy if you want to win this. You have to have a really smooth team where everyone knows what they're doing. It's like Formula One, just without the driver," said Polani.
"These robots do not have foot sensors. It's like having someone tie soles of wood to your feet."
While Chess was once considered a challenge for AI, it can now beat humans hands down. This, said Polani, was a bit of a let down as it didn't let people "learn anything about the nature of AI." But give AI a robotic body, make it interact with objects and other robots in space and time, and the challenge becomes significantly harder, as researchers must essentially coordinate teams of non-sentient beings in the physical world.
"There's not just one player but multiple players against multiple players, with everything happening in continuous time. This makes it much harder to plan and predict things," explained Polani.
In robot football, where your charges are pretty slow, strategy is key. Polani attributes his team's prowess last year to their software; they programmed their robot footballers to kick the ball to the side, rather than trailing around it.
"This buys us around ten to 15 seconds of repositioning. We were a bit slower than some teams at getting to the ball, but when we were there, we were accurate," said Polani.
While speed and accuracy are important, another challenge is keeping the robots upright. Most of the robots in the competition have no haptic feedback, so they have to process information from their surroundings visually.
"If you have feeling in the soles of your feet you can stay upright more easily. But these robots do not have foot sensors. It's like having someone tie soles of wood to your feet, or like [walking in] Dutch garden shoes," said Polani.
This year, Polani told me that his team were trialling out one robot with foot sensors. Given the unpredictability of these competitions, where whole robot teams can be felled by a faulty battery, a software or hardware blip, or simply overheating, Polani said he wasn't sure if they'd capitalise on the additional haptic feedback yet. But he expected other teams to follow suit next year.
Each year, the rules of the competition change and become harder. This year, both an AstroTurf floor (as opposed to carpet), and the need to program robots to recognise the colour white have upped the stakes—previously there were more visible red cues for the ball and yellow for goal posts.
Fouls are a pretty lax concept, and are often overlooked by the referee. A robot that stalls on pitch is dubbed an "incapable robot" and quickly lifted off the field to be resuscitated by its human handlers. Sometimes robots also end up bashing into one another, kicking or clinging to one another, but Polani said that they were disentangled and minor flaws ignored as the robots hadn't done it "intentionally."
"If a robot systematically destroyed a robot they would be disqualified—if that happens we're moving ahead," said Polani.
Since RoboCup's inception in 1997, where Honda unveiled a massive 130kg robot, things have come far. With technology becoming cheaper and more available each year, robots can be bought easily and smaller teams of researchers can get their robotic charges onto the field.
But both Polani and Noda agree that robots need a sense of "intention" if they're to be effective players in and out of the pitch. In other words, they need to be self-conscious and fully aware of their actions and desires.
Noda envisioned that neural network and deep learning technology could help in the future, but said that progress was still needed both in terms of hardware, motor, and sensor tech.
"We also need to establish a model of a human's thinking process and we need more sophisticated brain science, theory of social interactions, symbol grounding problems, and human learning theories," he said.
While no robot on the field looks like it can zip along at great speeds on the pitch, or work effectively without human intervention, both Noda and Polani are optimistic about seeing robot footballers on the real pitch by the 2050 World Cup.
"If you look at the progress that we've made since 1997, our robots are much more flexible and better-controlled," he said. Robot footballers in the 2050 World Cup, he added, were "implausible, but not impossible."