Abstract: The flexible terrain simulation system for a fully mechanized mining face serves as a test platform in a laboratory setting to simulate the intelligent coordination control of mining equipment in complex geological environments. One of its primary functions is to swiftly simulate the tilt state of the mining face floor in response to input curves. Given the platform’s numerous motion modules, strong coupling between moving components, and the intricate force dynamics during operation, achieving highly cooperative movement among these components poses significant challenges. To address these challenges, a training method for motion control strategies with mechanical perception capabilities has been proposed. This method constructs a structural reduced-order model through numerical simulation and deep learning. Combined with digital twin technology, a virtual-real integrated flexible experimental platform with mechanical sensing capabilities is established. By introducing a multi-agent deep reinforcement learning algorithm with a task allocation mechanism, mechanical information such as stress and deformation of each moving component is integrated into the multi-agent reinforcement learning environment. This enables each agent to perceive its mechanical state in real time and optimize task allocation based on training progress. The approach ensures safe cooperative motion while reducing the complexity of multi-agent learning, ultimately deriving optimal control strategies for the platform’s components. The results show that this method effectively reduces peak stress and deformation fluctuations in moving parts. It successfully coordinates 180 lifting cylinders to simulate complex floor tilt states, providing a robust test platform for intelligent control research in fully mechanized mining faces under challenging conditions.