AI Helps These Mini Humanoid Robots Learn to Play Soccer

Getting the hang of a sport like soccer can be tricky enough for people. It’s much harder for a robot. Bots typically aren’t great at adjusting to things on the fly, like moving to the ball or shooting around a defender. Despite that difficulty, Google’s AI research branch — DeepMind — recently held a one-on-one soccer match with mini humanoid robots that’s as impressive as it is fun.

While mini robots running around and kicking a ball may not seem all that complex, it’s a huge step forward for AI and robotics. That’s because, on top of playing pretty well, these bots taught themselves how to do it.

How These Mini Humanoid Robots Learn to Play

So, how does a mini humanoid robot learn to play soccer? The most obvious way would be to program them to follow a set sequence of moving and kicking. But that wouldn’t make for an entertaining match, and it’s not what DeepMind did.

Instead, researchers combined two machine learning techniques. The first is a method called reinforcement learning, where AI models learn through trial and error instead of humans telling them what’s right. The second is deep learning, which mimics the human brain through a complicated system of problem-solving steps.

This combination meant the mini humanoid robots could learn things with little to no instruction quickly. It also meant they could apply lessons they learned in simulations to the real world.

All in all, it took 240 hours of training for these bots to get the hang of playing soccer. Learning to play a sport in just 10 days for a robot with no basis of what sports are is an impressive feat.

What Soccer-Playing Robots Mean for the Future

As fun as robot soccer is, it’s not the most game-changing use for robots. Sports weren’t the ultimate goal of this research, either. The way these mini humanoid robots learned and how quickly they did it means big things for the future of AI and robotics.

AI That Learns at Lightning Speed

The main takeaway from DeepMind’s soccer bots is it shows a faster way to train complex AI models. Most machine learning algorithms can perform well in simulations pretty quickly.

However, applying those skills to the real world is often a challenge. People could train an AI robot in the real world instead, but that’s time-consuming. DeepMind’s solution strikes a better balance.

The soccer bots used reinforcement learning in simulations before using deep learning to learn to apply takeaways to real-world situations. That’s a powerful one-two punch of AI techniques — repeated trial and error in simulations is fast, while real-world applications are useful. Applying the same concept to other robots or AI models could have huge benefits.

Think of how companies claimed fully self-driving cars would be everywhere by 2021 but they still aren’t a thing. Real roads are unpredictable, so even cars that do well in simulations struggle with reality. The soccer robot approach could help them become a reality faster.

New Possibilities for Affordable Robots

This approach could also make AI-powered robots more affordable. While bots have been around for a while, they’re usually not cheap. AI features make them even more expensive because it takes a lot of time and money to get them working right.

Training AI models in less time means enterprises pour fewer resources into the product. As a result, they can sell the final version for less. Machine learning robots could become accessible to a much wider audience if that happens across the industry.

Affordable robots are great for consumers but even better for businesses. Smaller brands would be able to access these cutting-edge tools, and keep up with the speed and accuracy of their bigger competitors. Taking advantage of all of AI robots’ benefits would no longer be something only a few large corporations can afford.

Bots Redefining Versatility

DeepMind’s mini humanoid robots also paint a picture of a more flexible robot future. Some robots today are already fairly versatile — some have intelligent vision features to enable various games, for example — but many still rely on pre-programmed instructions. This new deep reinforcement learning approach could change that.

The combination of deep learning and reinforcement learning means robots can adapt to new situations or teach themselves new skills. Many of the soccer bots’ moves came from things they taught themselves, not from the programmers. That kind of adaptability means a single robot could perform a huge range of tasks.

Adaptable robotics would help organizations get more use out of one robot, helping make up for their upfront spending. It could also lead to more general-purpose robots for home or work use.

Pioneering Robot Applications

Fast-learning, adaptable robots could improve more specialized robot uses, too. Other advances in hardware let robots take on new forms. Flexible circuit boards, for example, cut space by 50% and weight by 90%, leading to smaller machines. Pairing this hardware with more adaptable AI could bring automation to new places.

Robots could learn to work in a busy warehouse alongside people if they can learn something as unpredictable and fast as soccer. Alternatively, they could help emergency responders navigate through wreckage to find and help survivors in a natural disaster.

These kinds of applications are difficult for many robots today, but the soccer bots prove how new AI techniques can assist technology in learning complex skills effectively. Taking advantage of this opportunity could have massive safety and cost benefits.

Why Mini Humanoid Robots Won’t Take Your Job Just Yet

Some people may look at these mini humanoid robots and get worried. After all, a robot that can learn something as human as soccer could perform other jobs, too. Job loss is already a concern with pre-programmed automation, and these AI-powered bots are significantly faster than scripted ones, so they’re even more beneficial for companies.

While these concerns deserve attention, an all-robot workforce won’t happen anytime soon. As flexible and impressive as these bots are, humans are still much better than robots at anything requiring adaptability or creative thinking. A robot may be able to play soccer now, but it’s not going to beat Lionel Messi in a match.

Even with these improvements, robots and people are best at different things. This project shows how robots and AI can become more reliable and accessible, but it doesn’t change that. Robots and humans are still best when they complement each other.

Mini Humanoid Robots Are a Glimpse Into the Future

As small as this use case is, DeepMind’s soccer-playing mini humanoid robots are an important step forward in technology. They show how AI can make automation more flexible, affordable and reliable in the real world. When more robots feature those kinds of benefits, both businesses and their customers will gain from them.