Identification appliquée à la robotique collaborative

Abstract

Robotics has played an important role on industrial changes over the last decades, allowing to reduce costs, improve quality, increase productivity and reduce dangers for employees. Even though robots were first designed to work in industrial scenarios inside cages, there has been a tendency over the last years to delete these barriers and allow the robots not only to share their workspace with humans, but also their tasks and objectives.This has widened their application area whose result is an increasingly frequent appearance of robots in our daily lives and a necessity to master the robot in a better way. Whether it is needed to design a control law, simulate the system to predict future states, or perform fault or collision detection, an accurate and reliable model is mandatory. The best way to do this is by gray-box modeling techniques: they make use of the knowledge of the physical laws that govern the system and of experimental data. The increasingly complexity of the systems and its surroundings, and the need to redefine the model while the system is running, bring these techniques back into the spotlight.In the context of this thesis, the focus is made on the development and enhancement of parameter estimation methods to be applied in the identification of the dynamic model of robotic systems. The case of collaborative robotics is studied by validating all the methods with the KUKA LBR iiwa 14 R820. However, our methods are general and applicable to a wide variety of robots, as results using the Pioneer LX 2-wheels differential drive mobile robot and the Stäubli TX40 industrial manipulator are also shown. The contributions are presented in two main axes. On the one hand, offline en-bloc estimation methods are treated, which are usually used to have a first idea of the model. Considering that the parameters have a physical meaning and that measurements are inherently noisy, the first proposed algorithm is robust against noise and ensures physical consistency of the estimates.The second proposed method addresses two other hindrances: the facts that manufacturers of commercial robots often hide important information and measurements to the users due to copyright and safety reasons; and the important but not fully understood friction phenomenon. The method identifies the model that the manufacturer has included in their controller in a reverse-engineering process, and uses it to estimate friction parameters.On the other hand, in systems that change during time or interact with an unknown dynamically changing environment, the estimation of the model and its parameters must be updated in an online basis. For this purpose, recursive variants must be analyzed. Aspects as algorithm initialization, stability of estimates and computing time gain importance. In this context, we first develop a new method that yields consistent estimates with noisy measurements, its robust against initial values and does not require an external simulation of the system. Second, although it is known that global identification is more accurate than sequential identification, sometimes the latter is the only possibility. We develop new methods which propagate the statistical distribution of estimates in the different identification steps, leading to no loss of information and more precise estimates. Finally, some of these methods are tested in several online scenarios which include human interaction and different payloads.