Predicting Part Orientation Distributions in Linear Feeders Using Simulation-Driven Deep Learning

Designing linear conveyor feeders with passive fences for automated part orientation remains largely trial-and-error because the final orientation distribution is difficult to predict reliably before physical testing. We present a simulation-driven deep learning pipeline that predicts the full distribution of final in-plane orientations for extruded, z-axis-symmetric parts interacting with linear feeders containing up to two straight or curved fences. Using Bullet physics-based simulation in CoppeliaSim, we generate 1,048 main part--feeder samples across 38 part geometries, plus 78 fence-generalization and 110 unseen-part samples for a total of 1,236 (41 unique parts), and train regression networks and a Variational Autoencoder, or VAE, to predict 360-bin orientation probability distributions. On known parts, the regression model achieves high accuracy on held-out test configurations, R² on circular CDFs = 0.97 ± 0.05, and on unseen fence combinations, R² on circular CDFs = 0.89 ± 0.11. Generalization to previously unseen part geometries is more challenging, with R² on circular CDFs = 0.75 ± 0.18, indicating that geometric representation and dataset diversity are primary limitations. We also evaluate VAE reconstruction on datasets generated from simulations at different iteration counts, 5--100% of 1000 iterations in 5% increments. While within-level reconstruction remains high, cross-convergence evaluation shows partial-iteration PMFs are far from fully converged labels in this dataset (overall CDF R² = 0.01 at 5%, 0.32 at 50%, and 0.87 at 75%), so reduced-iteration simulations do not substitute for full convergence here. Overall, the proposed approach provides a data-driven foundation for feeder analysis and design, with future work focusing on improved geometric generalization and physical validation for industrial deployment.