<p>Modular robots can adapt to various task scenarios and environments by rearranging their structural components and dimensions. However, their potential for versatility has not been fully explored in nonlaboratory environments, particularly on unstructured planetary terrains. This difficulty lies in the fact that the morphology and behavior of modular robots are highly intertwined with the terrain on which they stand. Achieving a concurrent design of robot configuration and motion strategy is essential to preserve the optimality of the reconfiguration schemes, as the completeness of the solution space can only be guaranteed if both are considered simultaneously. However, it is also challenging owing to the enormous joint candidate space. Existing research based on evolutionary algorithms, machine learning, or hybrid methods suffer from a range of limitations such as low goal-orientation and inadequate feature utilization. To this end, we incorporate a terrain-guided module and the spatio-temporal graph convolutional network architecture into the co-optimization framework to guide the optimization using agent features in both the spatial and temporal dimensions, which further accelerates the search and enhances the adaptability of modular robots. We conducted simulations using the Webots platform to validate our proposed method. Comparative studies showed that our framework produced reconfiguration schemes that exhibit highly efficient and appropriate morphology and behavioral adaptations toward several terrains.</p>