Label Me Maybe: A Unified Dataset for Workpiece Geometry, Discrete Poses, and Synthetic Rendering for Aerodynamic Part Feeding

High‐throughput part feeding remains a bottleneck for adaptable manufacturing: classic vibratory bowls achieve up to 200~parts/min but require workpiece‐specific tooling and manual tuning, limiting flexibility and inflating cost. We target an image‐based aerodynamic alternative in which parts slide down a V-shaped channel and are reoriented by timed air jets, guided by fast, discrete pose classification rather than full 6-DoF estimation due to the limited number of natural resting orientations. This paper introduces a design-grounded dataset and pipeline that link part geometry to slide-constrained poses. We contribute: (i) a standardized catalog of 39 parametrically defined workpieces built around four geometry axes (cross-section, length class, principal-axis symmetry, inside/outside features); (ii) a geometry-to-pose module that converts CAD into discrete, stable pose sets on the V-slide and exports canonical pose IDs with transforms and support polygons; and (iii) a pose-to-image data pipeline that renders synthetic images with light domain randomization and provides a small real validation set, all with consistent labels. This pipeline is developed with the intent of training and evaluating real-time pose classifiers, defining target states for physics-based simulation, and training reinforcement-learning policies for reorientation sequences. By standardizing workpiece definitions the dataset enables reproducible benchmarking and the pose and image generation reduces labeling overhead for industrial deployment.