DeepMind’s Self Improving Table Tennis Robots

Robots are already getting great at table tennis. In the future, they will improve on their own. Take DeepMind’s self improving table tennis robots for instance. It consists of a pair of robot arms that play table tennis against each other. When one learns a new way to attack, the other one adapts.

For the first step, a robot agent was trained using reinforcement learning in simulation to play cooperative rallies. It was fine-tuned for holding long rallies. The above video shows these robots in action.

[HT]