GNN-based deep reinforcement learning for computation task scheduling in autonomous multi-robot systems

Resource-constrained onboard robotic computers may encounter performance bottlenecks when handling computation-intensive tasks. In remote or extreme environments, external computing resources, such as cloud or edge computing platforms, are often inaccessible. Ensuring the timeliness of local computing tasks for individual robots, collaborative utilization of idle computing resources within an autonomous multi-robot system (AMRS) presents a promising solution. This study proposes a Graph Neural Network (GNN)-enhanced deep reinforcement learning (DRL) framework for the cooperative scheduling of multiple computing tasks within an AMRS. The GNN is employed to capture the structured matching relationships between task nodes and robot nodes, aggregating features through a message-passing mechanism. Given the dynamic variations in node activity within the AMRS, this study integrates a masking mechanism to filter out inactive nodes and utilizes a multi-head attention mechanism to distinguish the relative importance of neighboring nodes. The feature-enhanced node representations are then concatenated into variable-length sequences and fed into the DRL policy network for scheduling decisions. To further enhance scheduling accuracy, this study introduces a CNN–LSTM-based model that predicts actual task execution time by leveraging time-series data of robot resource utilization and task data volume, serving as a basis for scheduling decisions. The proposed framework is trained in a simulation environment constructed using real-world robotic data. Experimental results demonstrate that, compared to baseline algorithms, the proposed framework significantly improves the completion rate of cooperative computing tasks within an AMRS.