Image features for vision-based robot manipulation based on deep reinforcement learning

Deep Reinforcement Learning (DRL) provides a potential toolset that enables industrial robots to autonomously learn manipulation skills, but the learning efficiency (success rate within certain learning episodes) is the bottleneck. In this work, we ascertained well-designed environmental observations to be vital for improving efficiency. To determine the impacts of different observations, we conducted simulation experiments of robots grasping, and evaluated three popular categories of environment observations -positions of the Tool-Center-Point, raw images from a fixed viewpoint camera, and image features (Sobel, Laplacian, HOG, LBP). The results indicate “image features” proved to be superior to the others, they contribute to higher success rate and learning speed.