Contact stiffness estimation techniques for robotic manipulation

Abstract

This work addresses the problem of sti ness estimation for robotic tasks. Online sti ness estimation can be used to improve force tracking in explicit force control schemes being also important for realistic haptic feedback in telemanipulation tasks. It also enables accurate sti ness mapping and simulation of environment dynamics. Many applications involving contact can bene t from sti ness estimation. For instance, improved force tracking is useful to handle fragile organic tissues in robotic-assisted minimally invasive surgery. Enhanced haptic feedback allows the surgeon to have a better perception of contact forces, improving safety and allowing ner control. Environment sti ness estimation is useful in diagnosis, helping to detect pathologies through sti ness variations. Furthermore, accurate tissue simulation can be used in training and in task design. Applications of sti ness estimation, however, are not restricted to the medical area. These can be found in other contexts, such as industrial robotics. Examples include tasks involving contact, such as polishing or object assembly, that can bene t from improved force tracking. Also, inspection of object sti ness may be useful for quality control purposes. Many authors have described sti ness estimation as a complex problem. This complexity arises from multiple factors, where sensorial information with a poor signal-to-noise ratio is a major one. Uncertainty in the environment geometry is also relevant, as it obscures the relative positioning of the end-e ector w.r.t. the environment, creating dif- culties to detect free-space/contact transitions. This may be problematic, since most estimation techniques require accurate identi cation of the initial contact point. In this thesis, three new online sti ness estimation algorithms are developed for robotic interaction tasks. These algorithms completely prevent the dependency on environment position-based data, relying instead on forcebased data to obtain the sti ness estimation. Two di erent approaches have been adopted. The rst relies on the availability of explicit models of the control system and environment while the second uses implicit models trained with sensorial data. Two estimation algorithms, ASBA and COBA, were developed upon model-based techniques, while a third one, ANNE, relies on sensor-based techniques. ASBA is a sti ness estimator based on the active-state inspection of an active observer used in the control loop. The estimation is obtained by comparing the active state with theoretical predictions for di erent mismatches between nominal and real sti nesses. COBA is a sti ness estimator based on the analysis of the output of two force observers, each tuned with a di erent nominal parameter. The estimation is obtained by considering the force prediction error of both observers. Finally, ANNE is an estimator supported by arti cial neural networks (ANNs). It is composed of a layer of ANNs, whose outputs are combined by a fusion module that produces the nal estimation. Theoretical analysis, simulation and experimental results are provided, demonstrating that these techniques can e ectively achieve environment sti ness estimation. The algorithms described in this work present a contribution to online sti ness estimation in robotic tasks involving interactions with unstructured and unknown environments where contact geometrical data is unavailable or unreliable.