In the video work “Abend”, the contradiction between the uniqueness and reproducibility of the digital image is explored. Drawing on the logic described by Walter Benjamin, the digital copy ceases to be secondary in relation to the original and begins to function as an autonomous entity capable of transformation and degradation. Gradual distortions and the disintegration of visual form undermine the notion of stable identity, turning the self-portrait into a process rather than a fixed image.

The work also addresses the issues of digital embodiment and the posthuman subject, corresponding to the theories of N. Katherine Hayles, where the boundary between the body and its representation becomes permeable. The digital double exists as a separate yet connected form, whose “life” unfolds within a computational environment. The use of glitch aesthetics transforms technical failure into an expressive device, revealing the materiality of the digital medium. In this context, the work raises questions about authorship: between the artist, the system, and the algorithmic processes that shape the agent’s behavior and visual form.

The project was implemented in the Unity environment using the ML-Agents Toolkit, which enables reinforcement learning integration within a 3D simulation. Agents operate in a physically accurate environment based on NVIDIA PhysX, where movement is driven by a system of articulated rigid bodies (ragdoll). Behavior is trained using the Proximal Policy Optimization (PPO) algorithm, where the agent optimizes its policy based on observations, actions, and a reward function. The neural network policy model is implemented via PyTorch and trained in an external Python pipeline. Unity serves as the simulation environment for data generation, while training and weight updates occur outside the engine with subsequent synchronization. As a result, the system acquires a generative character, where agent behavior emerges rather than being directly prescribed.

The digital avatar was created using MetaHuman Creator within the Unreal Engine ecosystem based on user photographs. Through MetaHuman Identity, the face is reconstructed and matched to a parametric 3D model using computer vision methods. A highly detailed facial rig and PBR shading are then applied to achieve photorealism and precise animation. The outcome is a digital double suitable for real-time integration and animation.

2026