Abstract
Abstract In scaled-down sand table multi-agent formation experiments, existing control strategies often focus on path accuracy or formation stability, neglecting energy consumption constraints. This leads to a lack of energy awareness and inefficient energy utilization, and limited endurance under high-precision control. To address this issue, this paper constructs a continuous action decision layer driven by Eagle Strike SAC (Soft Actor-Critic) to achieve energy-aware control of speed and turning angle. It integrates the ant colony pheromone mechanism and GNN (Graph Neural Network) to complete group situational awareness modeling, and uses Wolf Pack FedAvg (Federated Averaging) to achieve distributed policy co-evolution, forming a cross-layer closed-loop architecture of “decision-perception-aggregation.” Simultaneously, a sand table physical energy consumption mapping model is introduced to embed actual power consumption into the learning process, improving formation stability while reducing energy consumption. Experimental results based on the Robotarium dataset demonstrate that the proposed method reduces energy consumption per unit distance by approximately 39% compared with MPC, while achieving a steady-state formation error of 0.03–0.05 m and a convergence time of 7.2 s. Compared with MADDPG and SAC baselines, the framework also shows consistent improvements in energy stability and trajectory smoothness. In addition, communication overhead is significantly reduced through selective parameter updates, without sacrificing convergence performance. These results validate that the proposed approach demonstrates a balanced and scalable solution for energy-aware multi-agent formation control, with improved energy efficiency under the evaluated settings.