This work presents a learning ap proach for industrial assembly tasks. lt was de veloped in the framework of the ESPRIT project B-Leam II. Often in industrial applications, often objects with a complex geometry have to be assembled. Compared to "top-down model-based" ap proaches, "data driven learning" approaches offer several advantages, such as (I) faster development ( 2) easier implementation and (3) computationally less expensive on-line. Model-based approaches be come very complex for real-world geometries and are ill-conditioned for relatively small objects. The test case consists of the insertion of an electric switch into a fixture. The learning task is to find the correct fixture position based on measured contact forces. Regression trees are compared with the cascade correlation architecture. Experiments with a KUKA industrial robot equipped with a force/torque sensor, validate the learning approach. Another contribution of this work is that the problem of tolerances is identified and assessed. The reported experiments show the effect of tolerances on learning performance.
Sensor Assisted Robotic Assembly, Learning Approach, Neural nets, Cascade Correlation Architecture, Regression Trees.
All articles published by Studies in Informatics and Control (SIC) are published in accordance with the terms and conditions of the Creative Commons Attribution License (CC BY-NC-SA 4.0).