Cooperative Motion Planning in Divided Environments via Congestion-Aware Deep Reinforcement Learning

In motion planning with partial observability, addressing uncertainty is crucial for preventing collisions and congestion, especially in the vicinity of constrained narrow areas connecting wider spaces, called hallways. In this work, we propose a cooperative motion planning algorithm that leverages congestion-aware deep reinforcement learning to alleviate collisions and congestion caused by uncertainty. Specifically, a relation analyzer is employed to build relational embeddings as agent representations, which are then fed into a subsequent motion generation network, enhancing the interpretation of the movements of other local agents. Additionally, a hallway map is constructed by merging the temporal arrival intents of neighboring agents, which is then used by a congestion-aware scheme to inform distributed motion planning. Simulations indicate that our algorithm outperforms the state-of-the-art in divided environments, producing better planning results in various scenarios. In summary, we present an adaptive and non-myopic distributed motion planning method in constrained scenarios and illustrate its performance in divided environments with various hallways.