Ball-cradling using reinforcement algorithms

To successfully learn and execute a new task autonomously is a complex problem for robots. Planning of this learning is based on human behaviors and human learning. For this purpose, a machine learning method, reinforcement learning has been developed. In this work, we modeled a problem called ball-cradling, where we made moving links as fingers, we taught the links using reinforcement algorithms to balance a ball without dropping, which falls from above. Q-learning, SARSA and Adaptive Heuristic Critic (AHC) algorithms were tested using Box2d simulator on this problem. Ball's position, ball's linear velocity, links' angle and links' angular velocity were used as state-space parameters. In the results, system managed to balance the ball without dropping it to the ground in single-link system for all algorithms. In two-link system, a successful learning hasn't been achieved. AHC algorithm showed a better learning performance compared to other algorithms.