We propose a novel framework for robotic metal scrap cutting in unstructured scrap yards. In this framework the robots and workers collaborate: the worker marks the cutting locations on the scrap metal with spray paint and the robot then generates the cutting trajectories. This leverages worker expertise, while deferring the dull, dirty, dangerous aspects to the robot. For the robot, this requires a 3-D exploration and curve reconstruction stage for path generation. We use a non-uniform rational basis spline (NURBS) model and a topological skeletonization method for path generation, and implement and compare these methods via simulations. These simulations employ a realistic sensor noise model and highly-detailed 3-D scans of complex, real-life scrap pieces. Real-robot experiments with three different shapes are also provided.