We recently talked on NCYT about Amazings about a humanoid robot that is capable of playing soccer. Now a different team of robotics is teaching the same skill to a quadruped robot.
For now, the robot isn’t skilled enough to take the ball away from a human opponent, but instead it is able to deftly maneuver a ball over terrain such as sand, gravel, mud, and snow, using an artificial intelligence technique to adapt. to the variable dynamics of the ball.
The work is carried out by a team made up of, among others, Gabe Margolis and Yandong Ji, from the Computer Science and Artificial Intelligence Laboratory (CSAIL), attached to the Massachusetts Institute of Technology (MIT) in the United States.
Perhaps the most fascinating aspect of this research and development project is that the roboticists have avoided directly programming all of the robot’s body movements, opting for a more automated approach. Through computer simulations, which include the robot, a ball, and terrain like those in the real world, they have made it the system itself that discovers the best way to handle the ball on each type of terrain.
Four thousand versions of the robot have been simulated in parallel in real time, which makes it possible to collect data 4,000 times faster than using a single physical robot.
The robot starts without knowing how to handle the ball. You only get a “reward” when you do it, or negative reinforcement when you mess up. Essentially, he’s trying to figure out what sequence of forces to apply with his paws. Over time, he learns to manipulate the ball better and better to achieve the desired speed.
The DribbleBot in action. (Photo: Mike Grimmett/MIT/CSAIL)
The robot can circulate without problems on terrain that it has never been on. If he falls, he is able to get up without help.
Although teaching a robot to play soccer seems frivolous, the real use of it is not in putting it to compete against Lionel Messi but in executing rescue missions for which a lot of expertise is needed, such as those that are usually necessary after a catastrophe such as an earthquake or flood. Traditional wheeled robots can be slowed or even unable to move over rubble. On the other hand, a robot with legs that is skilled enough to play soccer is sure to be able to walk quickly over rubble and find victims long before conventional robots can.
The DribbleBot team will publicly present the technical details of their robot at ICRA 2023 (IEEE Conference on Robotics and Automation), under the title “DribbleBot: Dynamic Legged Manipulation in the Wild”. (Fountain: NCYT by Amazings)