Solving the Wire Loop Game with a reinforcement-learning controller based on haptic feedback

Nowadays, many control systems rely on model-based approaches, which require an effort increasing with the complexity of the tackled issue. An alternative to these approaches is offered by Deep Learning algorithms, capable of autonomously learn from experience and requiring minimal prior knowledge of the problem at hand. In this paper we focus on the specific task of solving a point-to-point navigation mission across a generic path, namely the Wire Loop Game, while only relying on haptic feedback. A serial manipulator is employed to accomplish the task and the motion control is entrusted to an algorithm trained with Reinforcement Learning. Differently from widespread applications of vision sensors in this context, the task is completed by exploiting only data received from the force and torque sensor installed on the end-effector of the robot. Information on collisions are used to let the cobot correct its trajectory and properly traverse the path. The entire training of the neural network is performed in a real environment to avoid biases due to incorrect modeling and to further highlight the non-necessity of a model. Experiments results show very good success rate in completing different paths with a completion rate of 94% over 160 attempts on 8 different paths.