Double-Task Deep Q-Learning with Multiple Views

Abstract

Deep Reinforcement learning enables autonomous robots to learn large repertories of behavioral skill with minimal human intervention. However, the applications of direct deep reinforcement learning have been restricted. For complicated robotic systems, these limitations result from high dimensional action space, high freedom of robotic system and high correlation between images. In this paper we introduce a new definition of action space and propose a double-task deep Q-Network with multiple views (DMDQN) based on double-DQN and dueling-DQN. For extension, we define multi-task model for more complex jobs. Moreover data augment policy is applied, which includes auto-sampling and action-overturn. The exploration policy is formed when DMDQN and data augment are combined. For robotic system's steady exploration, we designed the safety constraints according to working condition. Our experiments show that our double-task DQN with multiple views performs better than the single-task and single-view model. Combining our DMDQN and data augment, the robotic system can reach the object in an exploration way.