Modeling of Load-dependent Friction in Robot Joints Using Long Short-term Memory Networks

The application of industrial robots (IR) in machining has many potential advantages such as flexibility and a large work-space. However, due to their functional structure, IR show weaknesses in absolute and path accuracy compared to machine tools. Model-based compensation techniques can be used as a solution for which precise modeling of the robot dynamics and its influences is required. Friction is responsible for a large portion of the total torque, especially at low speeds for which it shows a highly nonlinear behavior. Friction possesses many influencing variables that are not considered in simple analytical models, such as temperature or load. This paper addresses the latter in analyzing models that are able to model the influence of the load on the IR. For this purpose, analytical models are considered and validated using the first and second axis of an IR. In addition, Long Short-term Memory networks are analyzed as a data-driven technique in order to assess its ability to model highly nonlinear friction behavior, while being able to incorporate many input variables.