Comparative Assessment of Reinforcement Learning Algorithms in the Taskof Robotic Manipulation of Deformable Linear Objects

摘要

Reinforcement learning systems in robotics are still limited in their number of practical applications. They are often considered as unstable and difficult to implement. Moreover, very often, they demand a significant number of trials to the convergence, which may often be treated as a critical challenge in their application. However, gathering the data from the simulation can be the solution to that problem. In our paper, we are providing a comparative assessment of reinforcement learning algorithms in the task of robotic manipulation of Deformable Linear Objects (DLOs). We provide a comparison of four methods that work on the simulated robot. The tests were performed for two tasks - one is reaching, and the other is the folding of the DLO to the predefined, sinusoidal shape. The obtained results could be treated as a guideline for other researchers on the performance of RL methods in robotic manipulation tasks.