FLUX 3 Is Quietly a Robotics Engine: Black Forest Labs' Video Model Now Controls Robots at Audi

FLUX 3 Is Quietly a Robotics Engine: Black Forest Labs' Video Model Now Controls Robots at Audi

To generate a realistic video of a glass tumbling off a table, a model has to understand gravity, surface contact, weight, and the moment of impact. Black Forest Labs argues that this same understanding is exactly what a robot needs to catch the glass before it hits the floor.

That argument now has a production deployment behind it. FLUX 3, described by Black Forest Labs as a new multimodal foundation model, generates images, video, and audio from a single backbone. It also controls physical robots on an Audi production line through a partnership with mimic, a company that builds its own robots and brings expertise in robot learning, dexterous manipulation, and production deployment 1.

All of this comes from one source: that same July 23 blog post. Every quality and deployment claim below is self-reported by Black Forest Labs and mimic.

The thesis starts with compute. Video prediction accounts for over 95% of FLUX 3's total training cost 1. Generating realistic video is expensive because a model cannot fake physics. If it misunderstands how objects interact, the output looks wrong 1. Audio, by contrast, makes up less than 0.5% of the tokens in a 720p video with sound 1.

Robot actions, BFL argues, occupy a similar structural niche to audio: a low-dimensional signal tightly coupled to visual observations 1. Once a model has spent the compute to understand physical reality through video, predicting what a robot arm should do next is not a separate capability. It is the same world model decoded through a different output.

The test was whether adding actions would permanently degrade video quality. In a large-scale training run, BFL added action prediction to FLUX 3's curriculum and measured the impact. Human ratings on text-to-video and image-to-video tasks initially dropped by up to 10% 1. Within 3,500 training steps, the model recovered its full previous video quality while simultaneously predicting robot actions 1. BFL's interpretation: incorporating the action modality caused a temporary disturbance as the model realigned its internal representations, but no permanent capacity cost 1. These are the company's own human evaluation metrics, normalized to pre-action baselines, and no third party has verified them.

The result builds on work BFL calls Self-Flow, which the company says unifies generation quality and representation quality in a single training framework. In their internal testing, Self-Flow improved video generation quality and robot control success rates in simulation simultaneously, rather than trading one against the other 1. FLUX 3 is the scaled-up application of that approach 1.

The Audi deployment turns the experiment into a product. FLUX-mimic works by training a lightweight action decoder on top of intermediate features extracted from FLUX 3's video prediction path 1. The model's learned understanding of visual physics feeds directly into deciding what a robot should do next. mimic contributed its robots and full-stack deployment system; BFL contributed the foundation model and multimodal training expertise 1. FLUX 3 itself was trained on tens of millions of hours of general video plus hundreds of thousands of hours focused on human and robot manipulation tasks 1.

Companies that have poured billions into video generation models may have been building partial world models all along. If realistic video prediction genuinely requires understanding contact, weight, and causality, the line between generative AI and robotics AI is not a technology gap waiting to be bridged. It is a framing gap that a lightweight action decoder can exploit.

The evidence so far is one company's blog post, one partner's production line, and metrics nobody outside the partnership has checked. The thesis is testable. Action prediction either degrades video quality permanently or it does not. A single backbone either drives a robot arm or it cannot. BFL says it does, and points to Audi as proof. Independent verification would turn a well-argued claim into something the industry has to confront.