From Torque Estimation to Physical Human-Robot Interaction: the Role of Direction-Dependent Friction

Abstract

Accurate force regulation is critical in physical human-robot interaction (pHRI), where even small force overshoots can lead to discomfort or unsafe contact. This paper presents the identification and validation of symmetric and asymmetric linear friction models and investigates how these modeling choices affect interaction behavior. Both models improve torque prediction, with only moderate differences between them observed in joint-space level evaluation. However, comparable joint-space accuracy does not necessarily translate to comparable contact behavior. The symmetric model, despite achieving good torque estimation, results in larger force deviations and less consistent transient behavior. In contrast, the asymmetric model captures direction-dependent effects, resulting in more stable and consistent force responses. Experimental results show that, based on the average MAE across all joints, friction modeling reduces torque-prediction error by 44.5% with the symmetric model and 54.7% with the asymmetric model on the physical robot. In MuJoCo simulation, comparable reductions of 44.5% and 53.0% are observed, respectively. In contact experiments, both friction models reduce force overshoot and settling time, with the asymmetric model achieving the largest reduction in overshoot (22.7%).