Adaptive Feedforward Control Based on Real-Time Payload Estimation for Improved Tracking Performance

Abstract

Dynamics-based feedforward control has been shown to significantly improve the trajectory tracking accuracy of many kinds of industrial robots. However, the effectiveness of such schemes is limited by the fidelity of the underlying inversedynamics model, which in current implementations typically neglects joint friction and assumes a fixed, pre-specified payload mass. Payload and friction are inherently time-varying during operation: payload changes discretely at every gripping event, while joint friction varies with temperature, load, and velocity over the course of a production shift. This paper extends an existing Digital-Twin-based feedforward architecture in two ways. First, the payload mass, previously fixed as a static model parameter, is reformulated as a runtime input to the inversedynamics model and continuously updated using current data from the motor drives, enabling real-time payload estimation without additional sensors. Second, the architecture is combined with a recently developed sparse data-driven residual model that captures friction and other unmodeled nonlinearities. The combined adaptive feedforward controller is validated experimentally, demonstrating highly improved tracking accuracy across multiple payload configurations compared to both the pure feedback baseline and the non-adaptive feedforward controller.