Dynamic identification and control issues for collaborative robots: the KUKA lightweight robot LWR 4+ case

Abstract Accurate dynamic models are essential for high-performance control, both for trajectory tracking and physical human–robot interaction involving collision detection, reaction, and force control. This work presents a comprehensive framework for dynamic modeling and identification of the KUKA LWR 4+ collaborative robot based on two complementary approaches. The first exploits signals provided by the KUKA Fast Research Interface, combined with nonlinear joint-friction identification to improve torque prediction. The second is a general methodology applicable to collaborative robots with or without joint elasticity. We present the first complete identification and experimental validation of the KUKA LWR 4+ dynamic model using our Penalty-Based Parameter Retrieval framework, demonstrating its ability to retrieve a physically consistent model for a collaborative robot with elastic joints. For the first time for this robot, the approach yields a complete physically consistent set of dynamic parameters suitable for Newton–Euler algorithms, simulation, and internal wrench estimation. The resulting model is compared with the KUKA LWR 4+ model available in CoppeliaSim, which relies on empirically estimated CAD-based parameters. A procedure is also proposed to incorporate an unknown payload without complete re-identification. Although the resulting model accurately captures the robot dynamics, its direct use in torque control requires accounting for the low-level torque loop, which only partially compensates for joint elasticity. A novel recursive least-squares algorithm is therefore proposed to compensate for this effect. Finally, the validated model is employed in sensorless collision detection, isolation, and reaction experiments, demonstrating its suitability for real-time interaction control.