Building on this, the creative possibilities of AR are open. Animated models of real objects can be created that can be changed by the user by gripping and moving them. AR thus promises simplified operation, greater safety and easier learning when handling robots and machines.

The wide availability of collaborative lightweight robots enables the integration of robots into existing manufacturing processes. A distribution of tasks that utilizes the strengths of humans (dexterity, adaptability) and robots (strength, precision) is particularly suitable for small-batch production of products with many variants. One challenge is to coordinate the joint work of humans and robots. Traditionally, this is done by means of a predetermined plan. Creating this is time-consuming: Collisions must be avoided and the overall duration should be minimized.

This requires the movements of humans and robots to be estimated as accurately as possible and humans to be trained in the process. On the other hand, the potential of humans to react flexibly to deviations and gradually improve processes under real conditions is not fully exploited. Just as when working with a human, it would be desirable for the human-robot team to be able to familiarize itself with the process in order to be able to optimally complete the manufacture of a product. During this phase, both partners try out different approaches. They have to react flexibly to each other.

The Chair of Robotics at the University of Bayreuth is working on methods for recognizing the human approach and announcing the robot's movement. Figure 1 illustrates the interaction between humans and robots in a synthetic assembly task. The human can choose from a large number of variants as to how the individual components of the product are assembled. In each of the variants, the components are picked in a different order and placed in the robot's workspace at different times. For the robot, this means that it has to adapt flexibly to which assembly step is possible and sensible next. Playing through all variants is not an option, as the possible arrangements of just a few subtasks already generate too many variants. Instead, artificial intelligence and neural networks are used to learn a simplified model of how humans behave. Based on the model prediction, the robot estimates what the human will do next and adapts its work step accordingly.