Task-Oriented Reinforcement Learning with Interest State Representation

Current visual-based reinforcement learning for robotic manipulation is plagued by the problems related to goal transferability and anti-interference performance. The problems are mainly due to the pixel-dependence of state representations, which directly leads to the trained policy being highly correlated with the original observation data. To address above problems, a novel state representation framework named Task-Oriented Reinforcement Learning (TORL) is proposed by integrating Mask R-CNN with the original PPO-Clipped. In TORL, four cameras are used to capture multi-view observations from environment. Mask R-CNN is used to detect the objects in the environment, and the bounding boxes of the interest objects in each view are extracted as task features to define interest state representation and interest reward prediction. Four experiments are designed and carried out, the results demonstrate that TORL can improve goal transferability and anti-interference performance while ensuring the learning efficiency and stability.