Task Space End-Effector Contact Force Estimation for Robotic Manipulators Using a Bayesian Augmented Interacting Multiple Model-Based Disturbance Kalman Filter

Accurate estimation of contact forces acting on robotic manipulators is essential for safe and high-performance physical interaction. However, classical disturbance Kalman filter approaches are limited by fixed filter gains and single-model assumptions, making them less effective under dynamic and uncertain conditions. This paper presents a Bayesian Augmented Interacting Multiple Model Disturbance Kalman filter (BIMM-DKF), which combines a Bayesian Neural Network (BNN) for learning joint friction and modeling discrepancies with an Interacting Multiple Model (IMM) framework to adapt to multiple exogenous force dynamics modeled as a Markovian jump system. The predictive mean of the BNN is used as residual torque compensation, while its variance provides a state-dependent process noise covariance, enabling automatic, noise-aware gain scheduling. To mitigate computational complexity as the number of models increases, the estimator is formulated in task space. The proposed method was validated using a 1/8-scale mini hydraulic excavator with highly complex friction dynamics. Experiments were conducted under two different scenarios, and the proposed estimator demonstrated superior performance on fast-varying external force estimation compared to two baseline contact force estimators.