Abstract: To enhance the safety and real-time performance of underground gas inspection in coal mines and to overcome the low efficiency and high risk associated with manual inspection, a networked intelligent control approach integrating perception and decision-making was developed. Relying on multi-agent reinforcement learning (MARL), the framework constructs an embedded dual-module system that combines object recognition with cooperative path planning. The perception module utilizes the lightweight YOLO-Float detection network, which is optimized through structural pruning and 8-bit quantization to balance computational efficiency and accuracy. Furthermore, by integrating multi-scale feature fusion and attention mechanisms, the model achieves high-precision recognition of gas pipelines, valves, and dynamic obstacles. The decision-making layer introduced local optimal guidance, crossover mutation and information self-adaptive update mechanism based on the ant colony optimization algorithm, and formed an improved ant colony optimization (IACO) algorithm to complete the path planning. On the Gazebo three-dimensional mine simulation dataset, the constructed model was compared with the reference model. A 24-hour small-scale continuous inspection test was also conducted in the real mine environment. The test indicators covered detection accuracy, root mean square error (RMSE), path planning success rate, convergence generations, system failure rate, energy consumption, and data packet loss rate, etc. The simulation results show that after 50 iterations, the path fitness of the model reached 0.92, the root mean square error was 0.15, the detection accuracy was 88.3%, the average accuracy mean was 90.1%, and the training time with 800 frames of data was 1.8 seconds, and the processing delay was 3.2 seconds. Compared with the traditional model, it was shortened by 28% and 20% respectively. In the actual mine tests, the detection accuracy remained at 87.9%, the comprehensive failure rate was 1.8%, the communication interruption rate was 0.6%, the path planning failure rate was 0.7%, the average response delay was 2.8 seconds, the energy consumption was 14.6 Wh, and the data packet loss rate was 0.4%. Even under weak illumination and high humidity conditions, the detection accuracy still remained above 85%, and the path planning success rate was around 95%. The research results show that this method has achieved high-precision and low-latency autonomous gas inspection in coal mines.