Multisensor Data Fusion for Skill Transfer Systems

Rui Cortes~ ao Ralf Koeppe University of Coimbra German Aerospa e Center, (DLR) Ele tri al Engineering Department Institute of Roboti s and System Dynami s Institute of Systems and Roboti s, (ISR) P.O. Box 1116 3030 Coimbra 82230 Wessling Portugal Germany email: ortesao isr.u .pt email: Ralf.Koeppe dlr.de Abstra t The paper des ribes the design of a data fusion module for skill transfer purposes. The data fusion paradigm is addressed. It onsists of two independent modules for optimal fusion and ltering. A new interpretation of the Kalman lter equations is done, to a hieve a \model-free" equation apable of following arbitrary variables. A sto hasti approa h is used to tune the parameters of interest for a ertain task. The fusion algorithm here presented is global, and ould easily be extended to any arbitrary system. It was su essfully tested in the Institute of Roboti s and System Dynami s at the DLR [7℄. 1 Introdu tion The roboti industry is in reasing very fast, reating new hallenges and demands in many roboti appli ations. The available sensor te hnologies give the possibility to equip robots with a urate and robust sensors, enabling to understand better the surrounding world [5℄, [2℄. Intelligent algorithms are being developed, based on Fuzzy Logi , Arti ial Neural Networks or other emerging te hniques, permitting robots to learn, navigate, and plan in the real world [11℄. Some high-level skill transfer ar hite tures based in the human-robot interfa e are now being applied with some interesting results [8℄. These new approa hes, whi h mimi the behaviour of the human operator, have the potential to learn human-like tasks and are on eptually di erent from the lassi al model-based approa hes. They don't need an expli it mathemati al model des ribing the robot-world intera tions, redu ing the omputation overhead, and also the omplexity of the task exe ution. With this philosophy, the robot should learn human-like basi skills, as a starting point N Data Sources S2 S1 Mapping Mapping Mapping Optimal Fusion Optimal Filtering (LKF) Raw Data Output Vectors