Alexander Schubert - Generative Action Networks (GAN)
AI-based Choreography Generator
Research Project 2021
Funded by Take Care / Fonds Darstellende Künste

With "Generative Action Networks", a formal performative-choreographic grammar is to be developed which, due to its modular character, can be learned in the first step and subsequently generated with the help of artificial intelligence. The objective is to develop a text-based action description and then to program a matching AI software. Exemplary performances are to be developed and generated as research stages. The work is intended as a foundational research for more extensive performative pieces in the coming years.
The choreographic language to be developed will be composed of modules such as body part, posture, movement, transformation and modification. This formalization pursues the goal of realizing a representation of movement sequences that can be easily grasped symbolically and combinatorially. Thus, complex choreographies, arrangements or performances can be represented in a way that is easy to understand for a machine. The representation of physical actions should be done in categories. On the one hand, these can be concretely anatomical and explicitly descriptive in their execution - e.g. turning the left hand by 90°. However, more complex postures and sequences - as well as abstract and conceptual descriptions - are also possible. A spectrum of complementary categories should be developed so that they can be combined in the formal language. The description is thus not just a sequence of instructions, but consists of links between elements of different categories. The categories can be freely chosen and adapted according to the piece and context and are freely scalable, e.g. with respect to style and degree of abstraction. The language thus obtained allows a simple, concrete text-based notation of diverse performative sequences and choreographies.
The notation system to be developed is designed in such a way that the software to be programmed can easily be given existing sequences of actions as input. The AI can thus receive reference performances as training sets. These then form the basis for the automatic generation of new performances by the AI. This can happen in real time - or be output as a complete notation / script. Thus, similar or completely new motion sequences can be generated. This concept shall be developed as basic research for my own work and as an open access tool for other artists. In my work it will be the basis for performative parts in pieces that deal with artificial intelligence, technology, control and corporeality. Exploring the relationship between humans and machines in terms of autonomy, communication and interaction are examples of diverse resulting research fields.