Self-Supervised Vision-Driven Trajectory Planning for Robotic Deburring

In footwear manufacturing, robotic deburring presents a unique challenge due to high variability in product geometry and flexibility of materials. We propose a novel, fully self-supervised framework that enables robots to autonomously learn deburring trajectories directly from vision, without CAD models or human demonstration. This approach is inspired by human cognitive strategies, where vision and motor planning co-evolve. Our method comprises three key stages: (1) Learn to See, where visual networks extract a latent representation of relevant information of burrs from sole images; (2) Learn to Act, where, through reinforcement learning, an agent learns to map the latent features to 3D trajectories; and (3) Learn to Transfer, which uses Sim-to-Real methods including domain randomization and few-shot fine-tuning to ensure real-world applicability. We validate the framework in both simulation and physical experiments, achieving robust trajectory generation for diverse shoe sole geometries.