Boosting Machine Learning Techniques with Positional Encoding for Robot Collision Checking

Self-collision checking plays an important role in robot motion planning. In sampling-based motion planning methods, collision checking is performed multiple times, so this procedure should be accurate and fast. Collision checking is often addressed in robotics through the application of machine learning methods or a relatively straightforward multilayer perceptron within a low-dimensional feature space. In this paper, we focus on incorporating positional encoding, commonly utilized in computer graphics, into the input vector used in the binary classification task. We investigate the enhancement in classification accuracy by utilizing this technique in self-collision checking. Our findings indicate that positional encoding contributes to improved learning of high-frequency functions and provides a more accurate representation of higher-frequency details within the trained relation.