Mimic Robotics connects FLUX 3 to industrial robots at Audi

Why it matters

Black Forest Labs is using one costly video backbone for creative software and industrial robotics, while Mimic Robotics supplies the scarce action data and factory deployment layer.

Robin Rombach (@robrombach), co-founder and CEO of Black Forest Labs, and Mimic Robotics co-founders Stefan Weirich (@stefan_weirich) and Elvis Nava (@elvisnavah) have connected an early version of FLUX 3 to industrial robots. The resulting system, FLUX-mimic, is being tested and deployed with Audi on manipulation jobs involving components, seals, cables and other flexible materials, the companies said on July 23rd. (bfl.ai)

The collaboration puts Rombach's central bet about visual AI onto a factory floor. Black Forest Labs became known for image-generation models built by researchers associated with Latent Diffusion and Stable Diffusion. With FLUX 3, Rombach is arguing that the expensive work required to predict realistic video also produces internal representations of motion, contact, weight and cause and effect. Mimic Robotics is trying to translate those representations into robot movements. (bfl.ai)

Weirich and Nava bring the industrial half of that thesis. Mimic Robotics emerged from ETH Zurich's Soft Robotics Lab in 2024, where Weirich worked on biomimetic tendon-driven hands and Nava studied machine learning for robotic manipulation. Nava completed his PhD at ETH Zurich after working across the ETH AI Center, the Institute of Neuroinformatics and Robert Katzschmann's Soft Robotics Lab. Mimic Robotics builds dexterous hands and control software designed to attach to conventional industrial robot arms rather than requiring customers to adopt a full humanoid machine. (elvisnava.com)

One backbone for pixels and robot actions

Black Forest Labs described the technical thesis in its FLUX-mimic announcement. FLUX 3 is jointly trained on images, video and audio. Mimic Robotics then trains a lightweight action decoder on intermediate features from the model's video-prediction path, converting the model's representation of a likely future into executable chunks of robot motion. The decoder reads those latent features directly, so FLUX-mimic does not need to render a video before moving a robot. (bfl.ai)

Black Forest Labs says video prediction consumes over 95% of FLUX 3's training compute. Audio represents less than 0.5% of the tokens in a 720p video with sound, according to Black Forest Labs. The economics behind the architecture are clear: once Black Forest Labs has paid to train a model on vast quantities of video, action prediction could become another output of the same backbone instead of requiring a separate robotics foundation model. Those percentages and the underlying training costs have not been independently audited. (bfl.ai)

In one training run, Black Forest Labs said adding action prediction initially reduced human ratings for FLUX 3's text-to-video and image-to-video output by as much as 10%. The scores returned to their previous levels after 3,500 training steps, while the model retained its action-prediction capability. That result supports Black Forest Labs' claim that robotics does not permanently consume capacity needed for its creative products, although Black Forest Labs has not published enough detail for outsiders to reproduce the run. (bfl.ai)

The approach gives Rombach a route into physical AI without abandoning Black Forest Labs' existing media business. FLUX 3 is in early access, with Black Forest Labs planning separate rollouts for video and audio generation, image tools, action prediction through selected partners, and an open-weight multimodal backbone. FLUX-mimic is the first commercial test of how far the shared architecture can stretch. (bfl.ai)

Mimic supplies the scarce data

Mimic Robotics' contribution is the data and deployment system that a general video model lacks. In Mimic Robotics' account of the collaboration, Nava and his colleagues describe a three-level data strategy: large-scale human video at the base, demonstrations captured through Mimic Robotics' wearable hardware in the middle, and robot teleoperation and deployment data at the top. The last category is the most expensive and difficult to collect. (mimicrobotics.com)

That structure reflects Nava's research focus on extracting useful representations from pretrained models and reducing the amount of task-specific training required downstream. His earlier mimic-video work reported 10 times greater sample efficiency and twice the convergence speed of conventional vision-language-action architectures in the researchers' evaluations. FLUX-mimic applies the same method to a larger video backbone and Mimic Robotics' industrial dataset. (arxiv.org)

Mimic Robotics says it has collected data from over 100 factory use cases, spanning kitting, component insertion, assembly and the handling of deformable materials. Mimic Robotics also says FLUX-mimic can run locally on a single Nvidia RTX 5090, with the backbone producing its internal representation in less than 80 milliseconds and the full system reacting in 101 milliseconds. These are company measurements rather than independently reproduced benchmarks. (mimicrobotics.com)

Mimic Robotics reported a 95% completion rate across 20 autonomous trials of a soft-body kitting task. An adapted version of Physical Intelligence's pi0.5 reached 55%, while a task-specific flow-matching baseline reached 70%, according to Mimic Robotics. The comparison is useful as an early indicator, though Mimic Robotics selected the task, supplied the training mix and adapted the competing model. No independent evaluation accompanied the launch. (mimicrobotics.com)

Audi is the test, not yet the proof of scale

Audi gives the founders a serious industrial environment in which to test the claim. Mimic Robotics showed FLUX-mimic performing an Audi car-door assembly task and said the system is being used on real production use cases. Christoph Schneider of the Audi Production Lab said the robots had handled soft-body manipulation work that conventional robotics could not perform. (mimicrobotics.com)

The scope remains unclear. Neither announcement specified the number of stations, operating hours, completed production units or whether FLUX-mimic is running continuously on a live assembly line. Audi's Production Lab exists to test technologies under production-like conditions before deciding whether they can support stable, high-volume operation. The Audi relationship establishes access to factory problems and parts; it does not yet establish broad deployment. (audi.com)

For Weirich and Nava, the partnership arrives after Mimic Robotics raised a $16 million seed round led by Elaia and Speedinvest in 2025, with Founderful, 1st Kind, 10X Founders, 2100 Ventures and the Sequoia Scout Fund participating. The financing was intended to expand production of Mimic Robotics' hands and grow its engineering and field-deployment operations. (ai.ethz.ch)

Black Forest Labs has far greater financial backing for the compute-heavy side of the collaboration. In December 2025, Black Forest Labs raised a $300 million Series B at a $3.25 billion post-money valuation. Salesforce Ventures and AMP co-led the round, alongside investors including Andreessen Horowitz, Nvidia, General Catalyst, Temasek, Canva and Figma Ventures. (bfl.ai)

That capital lets Rombach train a video model at a scale a young robotics manufacturer could not easily finance. Mimic Robotics contributes the robot hardware, scarce manipulation data and access to factory workflows. The partnership divides the problem along the founders' existing strengths rather than asking either side to build the entire stack alone.

FLUX-mimic now has to prove that video pretraining produces repeatable savings when a manufacturer moves beyond selected demonstrations. Faster task adaptation would change the economics of automating work that varies by product, material or station. Audi's tests provide the first evidence from an industrial partner, while the missing deployment figures leave the scale of that evidence tightly bounded.