Predicting Cable Dynamics with Physical Attention Bias
Physical attention bias using cable geometry improves long-horizon prediction of unseen cable motion.
The paper studies learned simulators for cables, where most error occurs at self-contact or contact with the floor. It adds a physical attention bias, an additive term on attention logits, and compares arc-length distance, Euclidean distance, both distances on separate heads, and no bias. On unseen cables, arc-length bias reduces prediction error by 15% and more than halves segment-length drift when attention is the only long-range link. Assigning both distances across heads is best or near-best on every reported metric.
- Adds a learned physical bias to attention logits.
- Arc-length bias cuts unseen-cable error by about 15%.
- That bias more than halves drift in segment length.
- Using both distances on separate heads performs best.
From the original article
1 min read at huggingface.co
Learned simulators for deformable linear objects (DLOs) such as cables have to predict the motion of cables they were not trained on and stay stable over long rollouts. Most of their error occurs where the cable touches itself or the floor. Attention over all pairs of cable segments can represent contact between parts of the cable that are far apart along its length, but attention has no notion of geometry.
By Avihai Giuili, Rotem Atari, Avishai Sintov · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.