Orientation Prediction for Robotic Manipulation: Angle Encoding Strategies for Linear Regression

Accurately predicting object orientation from visual information is essential for effective robotic manipulation in smart manufacturing environments, yet standard single stage linear regression approaches struggle with the inherent discontinuity of angular data at the 0°/360° boundary. This paper investigates angle encoding strategies for objects within images to address this challenge, evaluating both trigonometric and vector based representations of sine and cosine components within shallow learning frameworks specifically, Support Vector Regression and Random Forest. Using a curated subset of the MetaGraspNet dataset focused on screwdrivers, we augment and validate orientation annotations through geometric analysis of segmentation masks. Comparative experiments demonstrate that these angular encodings substantially reduce angular prediction error, with Support Vector Regression models employing vector encoding achieving a mean absolute angular error of 5.01°. While all models exhibit increased error under severe occlusion, the results confirm that the proposed encoding strategies can accurately predict object orientation, offering a practical alternative to more complex multi-parameter grasp representations or multi-stage prediction models.