A group of researchers has developed a novel system to improve movement and precision in quadruped robots, demonstrated in the robot’s ability to shoot soccer balls.
The researchers; from the University of California, Berkeley (UC Berkeley), Université de Montréal and Mila, used the quadruped robot A1, developed by UnitreeRobotics in its tests. To enable soccer-shooting skills, the robotic system had to incorporate motion control and motion planning, granting the robot the ability to understand its surroundings and the trajectory needed to score a point.
Such a system is challenging as the robot not only has to determine how to kick the ball to achieve a goal, but also has to physically swing its leg without losing balance. The target of a deformable soccer ball also proves challenging, given its unpredictable interaction with the terrain.
The team used a reinforcement model to train the robot, with the robot receiving a reward each time it successfully completed a desired action. Ground frictions were randomized to train the robot to maintain control and accuracy in a variety of terrains and circumstances.
Using its newly developed system, the team found that when applied to a quadruped robot, precise results were seen almost every time, with the robot capable of shooting a deformable soccer ball to random targets with high precision.
Moving forward, the team said it will develop more nuanced and complex soccer skills, with the hope of one day creating a fleet of autonomous robots capable of engaging in an entire soccer match without human intervention.
The team also said the system could be deployed in robots used in activities that involve deformable items, such as clothes and ropes – for instance robots used in search and rescue missions.
.