Suitability analysis of machine learning algorithms: Processing three-dimensional spatial data for automated robot control

Abstract

Global competition, rapidly rearranging market requirements and shorter product life cycles are expressed in constantly changing environmental conditions, which further complicate the demands on the production process. Given smaller batch sizes in small to medium-sized companies, the importance of flexibly varying handling tasks, which must be implemented through a robot gripping system, increases. Standardized workflows are difficult to establish in undefined environments since the products to be handled vary strongly in orientation and position.

The work aims to determine whether artificial intelligence can be developed through the combination of a color camera including an infrared depth measurement, which enables industrial robots to interact with the environment. The following two research questions arise: 1. to what extent can the potentials of artificial intelligence and its success of the recent period be adapted for the application of a robot gripping process and 2. how this symbiosis effects the use of industrial applications. The combination of intelligently controlled robotics using artificial intelligence and the processing of data without serverdriven computing power at the end device form the basis of the investigation. The behavior of neural networks in scenarios with a small amount of data is the focus of the question. The realization of artificial intelligence is carried out in an iterative approach and the development process is available in written form.

The overall context of the approach is questioned via a suitability analysis to gain an understanding of possible applications and to name the limits of the system in the given scenario. With this approach, it can be examined which factors support the use of neural networks in the outlined context and whether they can be used successfully, despite of additional aggravating environmental influences.