Audi deploys Black Forest Labs AI to automate soft-body robotics

Black Forest Labs and Mimic Robotics have deployed FLUX-Mimic, a video-action model based on the FLUX 3 foundation, at Audi’s production facilities. The system enables robots to perform complex physical tasks, including the manipulation of soft materials and flexible components, which are typically difficult for conventional automation. Audi reports that the technology assists employees and expands flexible automation in production and logistics.

The deployment demonstrates a reaction time of 101ms, running on a single NVIDIA RTX 5090 GPU. Audi’s Production Lab, represented by Christoph Schneider, confirmed the deployment is solving complex soft-body manipulation work previously impossible with conventional robotics. The model leverages FLUX 3’s multimodal capabilities to understand world dynamics, allowing for high sample efficiency and rapid task adaptation.

FLUX-Mimic is built on the FLUX 3 backbone, which was jointly trained across images, video, and audio from the beginning. Video prediction accounts for over 95% of the total compute costs in FLUX 3 training, forcing the model to learn physical dynamics such as contact, motion, weight, and cause-and-effect. The model uses a lightweight action decoder on top of intermediate features extracted from FLUX 3’s video prediction path.

In testing, adding action prediction caused an initial 10% drop in video generation quality, which recovered after 3,500 training steps without permanent capacity loss. When finetuning the backbone with the action decoder, FLUX-Mimic achieved state-of-the-art success rates in manipulation tasks. The system demonstrated "self-recovery" capabilities, where robots corrected missed grasps and completed tasks without prior demonstration of the error scenario.

The model’s backbone can generate a world representation in less than 80ms, with the full system achieving 101ms latency through optimised inter-process and real-time chunking. This efficiency allows the system to run on consumer-grade hardware while maintaining the speed required for real-time industrial automation.

Audi’s Production Lab has integrated the technology to handle tasks such as kitting parts into structured trays, inserting electronic control units, and assembling components. The system’s ability to handle flexible parts addresses the economic challenges of variant diversity in premium production, where conventionally programmed robot cells are often too costly to re-engineer for each case.

Schneider noted that the partnership with Black Forest Labs and Mimic Robotics is essential in pushing the frontier of physical AI. The deployment validates these innovations in real-world production environments, offering a scalable solution for flexible automation that was previously unattainable with traditional robotic systems.