Abstract: On-orbit assembly based on autonomous rendezvous and docking of modular spacecraft is a promising technology for constructing large-scale space structures. Safely and efficiently implementing assembly missions requires solving the trajectory optimization problem for modular spacecraft amidst complex environments with obstacles. This paper proposes a spacecraft trajectory optimization method based on differentiable collision detection and efficient initialization for the pre-assembly phase. First, collision-avoidance bounding boxes for the assembly module and the assembled structures are constructed using a group of convex primitives, and the spatial regions occupied by these convex primitives are described by inequality constraints. The complex geometry collision detection problem is transformed into the computation of the generalized distance between the convex primitives contained within the bounding boxes, and sensitivity analysis is employed to obtain the derivatives of the generalized distance with respect to configuration parameters. Subsequently, the optimal trajectory for the modular spacecraft is solved based on the augmented Lagrangian iterative linear quadratic regulator (AL-iLQR). The dynamic system modulation (DSM) method is used to generate a high-quality initial guess for the optimization problem, effectively enhancing the algorithm’s stability and convergence efficiency. Finally, the efficacy of the proposed method is verified through numerical simulation examples.