Predicting Cable Dynamics with Physical Attention Bias
Adding geometry-aware bias to attention made cable simulators generalize better on unseen cables.
The paper tests attention biases based on arc-length distance, Euclidean distance, or both, while keeping the model and training setup fixed. Most errors occur at contact points, either cable self-contact or floor contact. Arc-length bias cut prediction error by 15% and more than halved segment-length drift when attention was the only long-range connector. Using both distances across separate attention heads was best or close to best on every reported metric. HF Daily Papers' note
The paper tests attention biases based on arc-length distance, Euclidean distance, or both, while keeping the model and training setup fixed. Most errors occur at contact points, either cable self-contact or floor contact. Arc-length bias cut prediction error by 15% and more than halved segment-length drift when attention was the only long-range connector. Using both distances across separate attention heads was best or close to best on every reported metric. HF Daily Papers' note
score 4