Current robotic solutions are able to manage specialized tasks, but they cannot perform intelligent actions which are based on experience. Autonomous robots that are able to succeed in complex environments like production plants need the ability to customize their capabilities. With the usage of artificial intelligence (AI) it is possible to train robot control policies without explicitly programming how to achieve desired goals. We introduce AI Motion Control (AIMC) a generic approach to develop control policies for diverse robots, environments and manipulation tasks. For safety reasons, but also to save investments and development time, motion control policies can first be trained in simulation and then transferred to real applications. This work uses the descriptive study I according to Blessing and Chakrabarti and is about the identification of this research gap. We combine latest motion control and reinforcement learning results and show the potential of AIMC for robotic technologies with industrial use cases.

The use of flexible and autonomous robotics systems is the solution for the automation task of the production and intra-logistics environments. This dynamic context requires the robot to be aware of its surroundings through the whole task, also after accomplishing the gripping action. We present an anomaly detection approach based on unsupervised learning and reconstruction fidelity of image data. We design our method to enhance the dynamic environment perception of robotics systems and apply it in a palletizing robot, in order to perceive and detect changes to its surrounding and process after the gripping step. Our proposed approach achieves the performance targeted by the considered industrial requirements.

The use of flexible and autonomous robotic systems is the solution for the automation in dynamic and unstructured industrial environments. This context requires the robot to be aware of its surroundings throughout the whole manipulation task, also after accomplishing the gripping action. This work introduces the deep post gripping perception framework, which includes the post gripping perception abilities realized with the help of deep learning techniques, especially unsupervised learning methods. These abilities help the robot to execute a stable and precise placing of the gripped items depending on the required task. We describe the development of the framework based on an established literature review. The result of the work is a modular design of the framework with the help of three functional components used to build planning, monitoring and verifying modules.

The increasing individualization of products reinforces the importance of decoupled factories in production processes. Artificial intelligence (AI) is a recognized technology for problem solving and accelerates automation by enabling systems to act independently. In the field of robotics, there are new deep learning approaches which make robotic control systems human independent. This work provides a literature overview of the current state of development methodologies, showing that there are only limited methods available for the development of artificial intelligent robots. We present a novel development methodology based on artificial intelligence, particularly deep reinforcement learning. The so-called Q-model can enable robots to learn specific tasks independently. In summary, we show how an AI-based methodology assists the development of autonomous robots along the product lifecycle.

Robotic systems need to achieve a certain level of process safety during the performance of the task and at the same time ensure compliance with safety criteria for the expected behaviour. To achieve this, the system must be aware of the risks related to the performance of the task in order to be able to take these into account accordingly. Once the safety aspects have been learned from the system, the task performance must no longer influence them. To achieve this, we present a concept for the design of a neural network that combines these characteristics. This enables the learning of safe behaviour and the fixation of it. The subsequent training of the task execution no longer influences safety and achieves targeted results in comparison to a conventional neural network.