Abstract: To address the issues of single-modal perception failure and insufficient coordination between control and perception in inspection robots under the complex underground coal mine environment, and to improve the accuracy and efficiency of autonomous inspection, a study on multi-modal perception collaborative control was carried out. Firstly, the environmental characteristics such as narrow and long roadways, high dust concentration, and strong electromagnetic interference in underground coal mines, as well as the core requirements including equipment condition monitoring and environmental parameter monitoring, were analyzed. An adaptive hardware architecture was designed, including a tracked mobile platform and a multi-modal perception module, and a multi-channel synchronous data acquisition and preprocessing system was constructed. A three-layer fusion framework covering the data input layer, feature extraction layer, and fusion decision layer was proposed. A timestamp alignment and dynamic weight fusion strategy was adopted, combined with an attention mechanism to achieve cross-modal feature-level fusion, optimizing feature dimensions and enhancing robustness. A perception-control collaboration strategy was designed, a task priority quantification mechanism based on urgency, influence scope, and duration was established, an improved A* algorithm introducing environmental risk values was proposed for motion planning, and a collaborative control algorithm linking perception feedback with actuators was constructed using a model predictive control framework. The experimental results show that the equipment fault recognition accuracy of the multi-modal fusion method reaches 95.7% (23.4% higher than that of the single-vision method); the gas concentration detection error is only 4.8%, and the obstacle missed detection rate is 1.2%. The improved A* algorithm has an average path length of 115.8 m, an obstacle avoidance success rate of 97.5%, and a planning time of 68 ms in complex obstacle scenarios, with the path shortened by 12.8 m and the obstacle avoidance success rate increased by 20.8% compared with the traditional A* algorithm. The collaborative control algorithm has a task switching time of 1.5 s, an abnormal handling accuracy of 96.7%, a target positioning error of ±5.3 cm, and an average energy consumption of 229 W in emergency response scenarios.