KinestheticGraspExperiment

Prev: Minimal Cognitive Behaviours	Up: Home	Next: Home

Table of contents

• Learning by demonstration

Learning by demonstration

The KinestheticGraspExperiment plugin implements an experiment in which an iCub robot learns to reach and grasp an object placed on a table. Unlike the GraspExperiment plugin, in which the robot is trained using a genetic algorithm, in this case both a supervised learning algorithm and a genetic algorithm are used. More specifically, the robot first acquires the ability to perform approximated reach and grasp through a learning by demonstration procedure and then its skills are refined by the application of evolutionary robotics techniques.

The iCub robot is placed in front of a table with a red ball on top of it, which can assume random positions. The learning procedure is made up of two phases. In the first one the robot arm is moved by external forces to reach and grasp the ball, as if a teacher was guiding it. During this movement, the sequence of postures is recorded and used to train the neural controller of the robot with the Levenberg-Marquardt algorithm (similar to the classical back-propagation algorithm, but with a much faster convergence time). Following this initial training, the robot undergoes an evolutionary process to fine-tune the parameters of the controller, in order to perform effective movements.

The experiment demonstrates how the combination of the two learning strategies allows to obtain robust solutions in a relatively short time. The evolutionary process, in fact, does not need to start from scratch: the behaviour at the end of the first learning phase, while rather rough, provides a starting point in the fitness landscape that is already near to a good solution, greatly speeding up the search process.

References

Valenti M. (2013). Learning of manipulation capabilities in a humanoid robot. University of Rome, La Sapienza: Master Thesis