DeepRL-Based Robot Local Motion Planning in Unknown Dynamic Indoor Environments

Robots are progressively being applied in different tasks, spanning from industrial applications to space exploration, enhancing human capabilities. Difficulties in performing many tasks such as navigation, target recognition and obstacle avoidance must be overcome. This work proposes a novel Deep Reinforcement Learning approach to solve robot motion planning in environments populated by both static and dynamic obstacles by exploiting a Dueling Double Deep Q-Network (D3QN) framework, with a costmap-based state representation, Prioritized Experience Replay (PER), reward propagation and curriculum/transfer learning. Evaluation was carried out in Gazebo simulation environments, and the presented results highlight the proposed framework’s performance in both static and dynamic scenarios.