Curiosity Driven Deep Reinforcement Learning for Motion Planning in Multi-agent Environment

Motion planning in dynamic multi-agent environment tends to be a challenging feat by itself. If the agent is also required to complete an elaborate task the complexity of the problem substantially increases. Heuristically designing policies for unstructured environments can be unfeasible and time consuming for certain scenarios. We propose a Deep Reinforcement Learning approach for a continuous multi-agent setting that is robust enough to handle high-level task accomplishment and low-level collision avoidance control. To alleviate disadvantages of a sparse reward environment we introduce intrinsic reward inspired by the curious behavior of animals and humans. As for the low-level, collision avoidance segment of control, we introduce a general reward function. Furthermore, we formulate an agent-centric state space and implement robust Reinforcement Learning algorithm that is capable of handling reward signal from both sources. Suitable 3D physical environment is constructed to examine the feasibility of our approach. Subsequently, case study is performed and effectiveness of our method is validated when compared with the state-of-the-art Reinforcement Learning algorithm.