Robots learned to play football thanks to machine learning

They did it well. Artificial intelligence specialists from DeepMind, a Google company, used machine learning to teach tiny robots to play football. They described the process of developing robots in the journal Science Robotics. DiscussRobots have learned to play football thanks to machine learning© Tech Xplore

As machine-learning robots become increasingly popular, computer engineers continue to look for new applications for such artificial intelligence tools. One of the tasks that has long attracted the attention of scientists and the general public is the creation of robots capable of performing traditionally human tasks that are difficult or tedious for humans.

The basic design for most such robots usually involved the use of direct programming or method imitation. In this new study, the team applied machine learning to the process and created tiny robots (about 510mm tall) that play football surprisingly well.

The process of creating the robots involved developing and teaching two basic reinforcement learning skills in computer simulations, such as getting up from the ground after falling or trying to score a goal. The system was then trained to play a full version of one-on-one soccer using a huge amount of video and other data.

Once the virtual robots were able to play as intended, the system was transferred to several Robotis OP3 robots. The team also added software that allowed the robots to learn and improve by first testing individual skills and then playing matches against each other on a small soccer field.

Watching their robots play, the researchers noted that many of the movements were performed more smoothly than robots trained using standard methods. They could get up from the field faster and more gracefully, for example.

The researchers say their AI robots played significantly better than robots trained by any other method so far.