Mysteric-Net: MIMO Hysteretic Friction-aware Lagrangian-based Network for Legged Robot

Abstract

Accurate dynamics modeling is crucial for achieving precise Ground Reaction Force (GRF) control and high-performance legged locomotion. However, real-world legged systems exhibit strong frictional effects with hysteresis and inter-joint coupling, which conventional static friction models or purely data-driven approaches often fail to capture. In this paper, we propose Mysteric-Net, a novel MIMO hysteretic friction-aware network that combines a Lagrangian-based formulation with a Temporal Convolutional Network (TCN). By embedding the physical laws of Lagrangian mechanics while modeling history-dependent frictional dissipation via the TCN, our framework accurately identifies the system dynamics, including complex friction and coupling effects. This paper demonstrates that the proposed method significantly improves the accuracy of inverse dynamics estimation on a robotic leg. Furthermore, this paper shows that the learned model enables the design of an effective feedforward controller that mitigates friction and enhances tracking performance over conventional baseline methods.