SPADE: Sketch-guided Path Planning Augmented with Diffusion Experts

Incorporating human preferences into path planning for Autonomous Mobile Robots typically requires complex reward engineering or costly teleoperation. Recent imitation-learning frameworks such as SKIPP let operators sketch desired paths, but they suffer from limited generalization to unseen environments and fragile data-collection tools. We introduce SPADE, a framework that addresses both problems through two contributions: (i) an open-source ROS 2-based annotation tool for robust demonstration collection, and (ii) a novel Conditional Diffusion-augmented Behavioral Cloning (Cond-DBC) training strategy. In Cond-DBC, an image-conditioned diffusion model, applied via FiLM layers, serves as an offline expert that guides a compact U-Net policy during training by providing a margin loss over the path channel. On a 22,000-instance dataset spanning ten maps, a medium Cond-DBC model (1.9M parameters) achieves 39.1% lower Absolute Pose Error and 33.5% lower Fréchet Inception Distance than the large SKIPP baseline (31M parameters), while preserving real-time, on-edge inference.