Sensorless Interaction Force Estimation via a Physical-Neural Adaptive Super-Twisting Momentum Observer

This paper proposes a physical-neural adaptive super-twisting momentum observer (ASTMO) to estimate interaction forces between a robot manipulator and uncertain environments. The proposed method reduces reliance on force/torque sensors while maintaining high estimation accuracy. Concretely, traditional dynamic modeling and identification methods often encounter challenges like unmodeled dynamics and parametric uncertainties, which may lead to significant modeling errors and adversely impact the accuracy of force estimation. To overcome these limitations, a hybrid model integrating physical dynamics with neural network corrections is proposed, introducing a unified framework that concurrently leverages model-based and data-driven strategies. The dynamic model is initially developed using the Newton-Euler method, followed by parameter identification to establish the physical model. A backpropagation (BP) neural network is then utilized to capture and correct residual model errors. Utilizing such a hybrid model, an ASTMO is constructed for external torque estimation. This effectively attenuates the chattering effect commonly associated with sliding mode control and improves estimation precision under uncertain operating conditions. Experimental validation on a six-degree-of-freedom (6-DoF) manipulator using a Beckhoff controller demonstrates the effectiveness in external torque estimation without a force/torque sensor.