Abstract

Accurately modeling friction in robotics remains a core challenge, asrobotics simulators like Mujoco and PyBullet use simplified friction models orheuristics to balance computational efficiency with accuracy, where thesesimplifications and approximations can lead to substantial differences betweensimulated and physical performance. In this paper, we present aphysics-informed friction estimation framework that enables the integration ofwell-established friction models with learnable components-requiring onlyminimal, generic measurement data. Our approach enforces physical consistencyyet retains the flexibility to adapt to real-world complexities. Wedemonstrate, on an underactuated and nonlinear system, that the learnedfriction models, trained solely on small and noisy datasets, accuratelysimulate dynamic friction properties and reduce the sim-to-real gap. Crucially,we show that our approach enables the learned models to be transferable tosystems they are not trained on. This ability to generalize across multiplesystems streamlines friction modeling for complex, underactuated tasks,offering a scalable and interpretable path toward bridging the sim-to-real gapin robotics and control.