Synergy-based policy improvement with path integrals for anthropomorphic hands

In this work, a synergy-based reinforcement learning algorithm has been developed to confer autonomous grasping capabilities to anthropomorphic hands. In the presence of high degrees of freedom, classical machine learning techniques require a number of iterations that increases with the size of the problem, thus convergence of the solution is not ensured. The use of postural synergies determines dimensionality reduction of the search space and allows recent learning techniques, such as Policy Improvement with Path Integrals, to become easily applicable. A key point is the adoption of a suitable reward function representing the goal of the task and ensuring one-step performance evaluation. Force-closure quality of the grasp in the synergies subspace has been chosen as a cost function for performance evaluation. The experiments conducted on the SCHUNK 5-Finger Hand demonstrate the effectiveness of the algorithm showing skills comparable to human capabilities in learning new grasps and in performing a wide variety from power to high precision grasps of very small objects.