Abstract: The complexity of soft-terrain environments in open-pit mines poses severe challenges to multi-agent reinforcement learning (MARL). Traditional geometric obstacle-avoidance strategies neglect the underlying wheel-terrain interaction mechanics. Meanwhile, existing MARL algorithms typically rely on soft penalty mechanisms and suffer from environment non-stationarity induced by joint policy updates, leading to poor safety and coordination performance on soft terrains. To address these issues, we propose a physics-informed reinforcement learning framework to learn cooperative truck-planning policies that strictly satisfy physical safety constraints while fully leveraging global environmental contexts. Specifically, the Bekker-Wong terramechanics model is embedded into the underlying formulation of a constrained Markov game, explicitly accounting for the bearing capacity limits and physical safety boundaries of the soft subgrade during policy optimization. Furthermore, a heterogeneous dual-stream perception architecture integrating multi-head self-attention and Convolutional Neural Networks (CNNs) is developed, enabling the agents to simultaneously process high-dimensional spatial physical fields and variable-length temporal-topological information from multi-truck interaction graphs. Additionally, we design a sequential policy-update scheme equipped with trust-region dual projection and dynamic priority assignment, which effectively mitigates the non-stationarity and deadlock phenomena caused by concurrent multi-agent updates. Simulation results in open-pit multi-truck coordination scenarios demonstrate that the proposed method consistently outperforms existing constrained MARL baselines in terms of physical safety, convergence speed, and spatial uniformity of cumulative traffic-load distribution.