MIT

DribbleBot is a very sophisticated four-legged robot that is made at MIT’s CSAIL and is capable of dribbling a soccer ball over rugged and varied terrains such as sand, grass, mud, and snow. Ahead of typical legged robots that are liable to fall over over rough terrains, DribbleBot, through reinforcement learning, is able to learn to change its leg motion over rough terrain such that it is able to mimic a person’s soccer ball dribbling motion. The robot, through simulations, is trained through learning how to move the ball through receiving a reward when it is able to successfully dribble and a penalty when it makes a mistake, a skill that keeps improving every day.

The robot is equipped with cameras and sensors through which it is able to sense and react to its world in real time. Sensors not only assist DribbleBot in learning its location but are also able to execute related moves to navigate through complex worlds. The robot is able to bounce back to its balance and get back to activity when it is down while it is dribbling a basketball because it contains a recovery mechanism.

The design of DribbleBot is part of a larger project of designing legged robots that are able to operate in a wide variety of off-road terrains, which is also preferred in applications such as search and rescue during disaster response operations. The robot performs very well in dynamic environments but still remains to be improved upon, particularly when navigating through hazards such as stairs and sloping terrains. But notwithstanding these areas of weakness, its design presents very good learning in combining locomotion and grasping and manipulation of objects and opens a direction towards designs of much more independent and agile robots in real-world applications.

By:
Vraj Parikh