Autonomous vehicles excel in high-risk missions due to their adaptability and cost-effectiveness. Trajectory planning, crucial in autonomous systems, designs motion paths based on task objectives and constraints. Traditional navigation relies heavily on pre-acquired maps and prior knowledge, struggling in unknown and dynamic environments. While end-to-end methods enhance the efficiency of trajectory planning, they often lack generalization and require extensive labeled data. We propose an innovative Instructional Learning Approach (ILA) method, utilizing differentiable cost mappings during training for implicit supervision, reducing the need for labeled data. To generate smooth, collision-free paths, this method integrates a Bi-Level Optimization (BLO) process and connects network updates with trajectory optimization. Our approach backpropagates task-level costs, enabling direct gradient descent updates to the network. Experiments show our method improves planning speed by approximately four times compared to traditional methods and demonstrates strong robustness to localization noise. The ILA method can effectively generalize to unknown environments and has better learning performance compared to baseline learning methods.