The shifting nature of obstacles and overall environmental conditions pose significant problems to robotic path planning within constantly dynamic surroundings. As many classical path-finding algorithms implement a heuristic for real-time processing, they fail to adapt correctly to provide optimal navigation solutions. It presents an innovative solution that applies the principles of Graph Reinforcement Learning (Graph RL) for optimal robotic paths to efficiently spatially model dynamically changing relationships and environments. A robot's workspace can be conceptualized as a graph structure where nodes and edges denote discrete robot states (or positions) that describe transitions that can be made from one state to the other. Features of graph nodes and edges are defined by mathematically modeled dynamic nature attributes like moving obstacles and time varying costs of the environment. This incorporation allows for representing complex spatial-temporal dependencies related to robust navigation in real- world scenarios. Our framework enables the robot agent to dynamically choose accurate steps to process its journey using GNNs for the structured input graph. Information from the entire neighborhood where the robot is situated is aggregated through GNN, allowing the trained agent to make strategically sound decisions. For the proposed approach of path choosing, we apply different model centers to allow the smart adaption of multi-direction decisions to include environments according to action defined reward structures dynamically. With the proposed solution, we utilize dynamic graph RL algorithms for streamlined learning with guaranteed stability. Findings from simulations with varying parameters show that our Graph RL model surpasses classical path planning methods and non-graph-based RL methods. The model displays superior performance, including shorter path lengths, lower collision rates, and faster environmental change adaptation. Further studies demonstrate how the graph-based structure is fundamental to performance improvement in learning. This research shows the benefits of applying graph-based deep learning to reinforcement learning in robotic navigation problems. This will also be useful in multi-agent systems in real-life applications, setting the stage for advanced autonomous robots that function in unpredictable environments.
Snousi et al. (Mon,) studied this question.