Driving · 2026-10-09

Predicting Cable Dynamics with Physical Attention Bias

Avihai Giuili, Rotem Atari, Avishai Sintov, Maya Bechler-Speicher

arXiv:2610.11975PDFCode

AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.

TL;DR

The authors study learned simulators for deformable linear objects (cables) and ask what geometry should bias global attention between distant cable segments. They add a physical attention bias to attention logits using either arc-length (intrinsic) distances, Euclidean (extrinsic) distances, or both on disjoint attention heads. Physical bias improves prediction on unseen cables; arc-length yields the largest gain when attention is the only long-range coupling. Code: https://github.com/avihaig/dlogps.

Why it matters

For cable dynamics, contact and elastic forces depend on different pairwise distances: arc-length for elastic forces and Euclidean distance for self-contact. Since attention over all pairs can model long-range interactions but lacks geometry, adding a geometric prior to attention can improve generalization to unseen cables and reduce rollout drift, especially when the attention stream provides the primary long-range coupling.

Method

  • Start from a hybrid graph transformer with two parallel streams: a GatedGCN on chain edges and dense multi-head attention over all vertex pairs; train and evaluate with local stream on/off and autoregressive rollouts.
  • Add a physical attention bias as an additive term on attention logits with a learned per-head decay rate; test variants using arc-length distance, Euclidean distance, or both on disjoint sets of heads (plus unbiased heads).
  • Train in MuJoCo on one-step prediction with random-walk velocity noise and evaluate on unseen cable instances across 400-frame rollouts using rel-l2 error plus physical violations (link-length drift and self-intersection rate).

Limitation

The authors note that the local-on ranking lies inside the window spread, test cables lie inside the training range, and a lowest mean does not by itself show that heads specialize.

Abstract (from arXiv)

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. A cable has two pairwise distances, which agree only while it is straight: the arc-length distance along the cable, which governs elastic forces, and the Euclidean distance in space, which governs contact. We add a physical attention bias, an additive term on the attention logits with a learned rate, and ask which distance it should use. We compare no bias, each distance alone, and both distances on disjoint sets of heads, keeping the rest of the model and the training protocol fixed. A physical bias improves prediction on unseen cables. The gain is largest when attention is the only mechanism that connects distant segments: there, the arc-length bias reduces prediction error by 15% and more than halves the drift in segment length. The Euclidean bias alone stays close to unbiased attention, while assigning both distances across heads is best or near-best on every metric we report. Code and per-run records: https://github.com/avihaig/dlogps.

Related papers

Publication record

Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.

First public release: 2026-10-08 · Indexed by Pub-AI: 2026-10-09 UTC

preprint

Predicting Cable Dynamics with Physical Attention Bias

Avihai Giuili, Rotem Atari, Avishai Sintov, Maya Bechler-Speicher

Preprint posted: 2026-10-08 · Original record ↗

arxiv
2610.11975
Explicit source versions (1)
  • v1 · 2026-10-08
  1. Preprint available

    Source date: 2026-10-08 · First detected: 2026-10-25 UTC · Historical baseline

    Source evidence (1)
Abstract · arxiv

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. A cable has two pairwise distances, which agree only while it is straight: the arc-length distance along the cable, which governs elastic forces, and the Euclidean distance in space, which governs contact. We add a physical attention bias, an additive term on the attention logits with a learned rate, and ask which distance it should use. We compare no bias, each distance alone, and both distances on disjoint sets of heads, keeping the rest of the model and the training protocol fixed. A physical bias improves prediction on unseen cables. The gain is largest when attention is the only mechanism that connects distant segments: there, the arc-length bias reduces prediction error by 15% and more than halves the drift in segment length. The Euclidean bias alone stays close to unbiased attention, while assigning both distances across heads is best or near-best on every metric we report. Code and per-run records: https://github.com/avihaig/dlogps.