Soccer-playing robots teach themselves to score

From Science Magazine.

Developing athletic ability requires practice, even for robots. Using an approach driven by trial and error called deep reinforcement learning, researchers at DeepMind helped these soccer-playing robots develop skills, agility, and techniques to improve their play–all at a higher level than could be manually programmed. The computer “agents” controlling the robots picked up the rules and nuances of the game by observing human soccer players, then playing against each other in both simulations and using real-life robots. Robots play each other to improve their game, get computational “coaching,” and even mimic defensive moves of human players. While these athletic machines may not be making starting lineups anytime soon, computer scientists hope that this ability to learn and perform complex tasks in unpredictable environments will prove useful well beyond the soccer field.

Read the research paper here: https://www.science.org/doi/10.1126/scirobotics.adi8022