Researchers Dr. Peter Kormushev, Dr. Sylvain Calinon, Dr. Ryo Saegusa and Professor Giorgio Metta at the Italian Institute of Technology, have created a learning algorithm called Augmented Reward Chained Regression (ARCHER), enabling iCub a 53-DOF humanoid robot to learn archery.
The ARCHER algorithm was designed exclusively for those problems in which the purpose of the mission has to be known in advance. For instance, prior information about archery is pre-determined which is hitting the center circle. Once it is achieved, the learning algorithm adopts a chained local regression method that continuously determines new policy parameters based on previous feedbacks, leading to a greater possibility of attaining the mission. In contrast to other learning algorithms, ARCHER utilizes a more affluent feedback data regarding the outcome of a rollout.
In order to teach the skill to the iCub robot, researchers utilized the learning algorithm to alter and synchronize the hands movement which was facilitated by an inverse kinematics controller. At the end of each rollout, the image processing unit identified the location where the arrow had hit the target by the Gaussian Mixture Models for color-based detection of the target and the arrow’s tip and this feedback is transmitted to the ARCHER algorithm.
It was reported that the iCub robot is 104 cm tall and the target was placed at 3.5 meters away from the robot. This study will be demonstrated at the Humanoids 2010 conference which will be held in the U.S in December 2010.