Google DeepMind, the AI arm of Google, has taken a unique approach to training robots by teaching them to play football. In a recent study published in the journal Science, the company explained how they used deep reinforcement learning to train bipedal robots to exhibit a range of agile behaviors on the football field.
The researchers at Google DeepMind created a framework that allowed the robots to play one-on-one soccer, where they developed new behaviors such as dynamic movements to recover from falls and strategic tactics like defending the ball. The robots trained using this framework were able to move faster and more efficiently than those with traditional scripted controllers.
The team used the MuJoCo physics engine to simulate the training of the AI agents before transferring the learned skills to small humanoid robots with 20 actuated joints. This approach helped the robots achieve faster walking, turning, kicking, and standing, as well as the ability to combine these movements to score goals, anticipate the ball’s trajectory, and block opponent shots.
“Robotics has long used games like soccer to test complex motion, coordination, and environment awareness,” Google DeepMind said in a post. The success of this training method suggests that the framework may be useful for even more complex interactions between multiple robots in the future.
Overall, Google’s innovative approach to training robots for football demonstrates the potential of AI in imparting new skills and behaviors to machines, paving the way for more advanced applications in various industries.