Physics-Regularized Learning for Robust Sim-to-Real Tactile Friction Estimation

Tactile friction estimators trained on limited regimes often fail under physically inconsistent or out-of-training contact dynamics. We present a physics-regularized variational autoencoder (VAE) that embeds Coulomb's friction law as a differentiable constraint, encouraging force-consistent friction estimates under physically inconsistent tactile observations. On 1,730 real DIGIT sequences, the frame-based model reduces zero-shot friction MAE by 30.5% relative to a standard VAE; on an extended 5,500-sequence set across flat, sharp, and hemisphere contacts, it retains a 25.5% advantage. The frame-based study isolates the local physics prior, while a matched GRU extension demonstrates compatibility with temporal context and provides the best zero-shot transfer performance among the evaluated configurations. Fine-tuning reveals a sharp trade-off: unconstrained models fit labeled real data far better, whereas the current physics regularization preserves a conservative Coulomb margin but is too restrictive to match baseline accuracy in label-rich adaptation. These results show that domain-specific physical constraints improve zero-shot and label-scarce sim-to-real tactile friction estimation while exposing limitations in instrumented consistency assessment and supervised adaptation.