Path planning is a crucial function for intelligent agents, with applications ranging from navigation to object tracking and rearrangement. Existing methods primarily focus on tasks with explicit goal and constraint specifications, limiting their applicability in scenarios where such specifications are difficult to define. This work introduces a novel data-driven paradigm called example-driven planning, which allows users to specify goals and constraints through sets of target and support examples, respectively. Formulated as a constraint optimisation problem, we propose a learning-based framework, namely DualGF, to address this task. Specifically, DualGF models the planning process using two gradient fields: a target gradient field, which attracts the agent towards the target examples, and a support gradient field, which repels the agent away from the support examples. Both fields are parameterized by graph neural networks and are trained using a denoising score-matching objective based on the provided examples. To strike a balance between these two fields, we introduce a gradient mixer that dynamically adjusts the mixing rate during execution. Experiments across four tasks (navigation, tracking, particle rearrangement, and room rearrangement) demonstrate our method’s scalability, generality, and effectiveness.