Researchers at Johns Hopkins University are successfully using positive reinforcement procedures to help robots learn. Reinforcers in the form of points are given to the robots upon the successful completion of a block stacking task. If the robot makes a mistake, no reinforcers are given.
Though the term isn’t mentioned in the video below, the team effectively uses differential reinforcement procedures to accelerate learning that would take a month, down to two days. They stress that the block sorting task is used for convenience, but the methods can be generalized to a myriad of other applications such as self-driving cars and surgical robotics.
For more on their work, and to meet the team, check out the video below.