MIT

DribbleBot, an innovative four-legged robot designed at CSAIL at MIT which has the ability to dribble a soccer ball on varying and difficult surfaces, sand, grass, mud, and snow, etc. In a way different from conventional legged robots, which are fragile on rough terrain, DribbleBot employs reinforcement learning to adjust the leg movements on textured ground, sort of the way humans roll soccer. Using simulated environments, the robot trains to control the ball by getting rewards when rolling the ball properly and facing a penalty when the ball roll is wrong, leading to an improvement of the robot’s skill over time.

The robot’s system is based on sensors and cameras on board which help it to perceive and adjust its environment in real-time. These type of sensors allow DribbleBot not only knowing where it is, but also makes the appropriate maneuvers in order to move in an intricate environment. While the bot is dribbling the ball, the bot also has a recovery mechanism that allows it to stand on its feet after it drops to the ground and that guarantees it can finish the task even if it falls.

The development of DribbleBot is a fragment of a wider work aimed at building legged robots able to operate under various (off-road) terrain, which are crucial for applications as search and rescue in disaster areas, etc. Although the robot is excellent in the dynamic environment, it still needs to be further improved, such as more robust way to deal with obstacles, e.g., stairs and slopes. However, the study does provide useful information both regarding the combination of locomotion and object manipulation and how to combine them, which brings potential to even more agile and autonomous robots in use-able situations.