An International team of ML scientists teaches simulated mannequins to play football. Science Robotics journal published the results, including the endearing videos of the AI mannequins taking their newfound skills into the pitch.  

Before the AI grudge match, the 3D AI humanoids were trained in some rudimentary locomotion. Then, the team of computer scientists led by Siqi Liu at DeepMind outlined a three-step ML framework designed to teach virtual AI humanoids a wide range of basic skills and motor functions. 

Initially, the AI mannequins watched videos of moving humans to learn through imitation. Next, a reward-based reinforcement learning approach taught the AI, mid-level skills such as kicking a ball. 

Finally, they gained high-level skills, including teamwork, and body control, by working together in teams under an advanced version of reward-based reinforcement learning, also known as “match”. 

No prior knowledge of football, its rules or objectives was fed into the ML algorithm. Yet the virtual players learned all these skills and played a tensely fought game against each other exhibiting collaboration, body control and decent technique.  

Among the familiar football skills mastered by the virtual players were jostling for position, clearing the ball away from the goal they are defending, rapid turns, through balls, kicks of varying height, slowing down before changing direction with the ball, and tackling and running into a defensive position.  

The simulated players achieved integrated control in a setting where movement skills and high-level goal-directed behavior were tightly coupled- a setting that is reflective of many challenges faced by animals and humans and where solutions would be extremely difficult to handcraft. According to the researchers, the research into teaching AI how to move and interact as animals or humans may help in the development of more animal-like robots.