Introducing FLUX-mimic, a next-generation Video-Action Model for general purpose dexterity, developed in partnership with Black Forest Labs. Late last year we published mimic-video and introduced Video-Action Models (VAM): a new family of robotics foundation models built on top of video generation models. We showed that robot control reduces to... visual prediction, and that robot capability is downstream of improvements in video modeling accuracy. The obvious implication was that advances in the video modeling frontier would directly translate to increased capabilities in end-to-end robot learning. FLUX-mimic is that thesis at frontier scale: We've applied our VAM architecture to the strongest video backbone available today, FLUX 3 from Black Forest Labs, and trained it on data from our own robots and wearables. General-purpose dexterity, running on a single GPU on premises. Because the model already understands world dynamics, it needs far fewer demonstrations to learn a new task. This is game-changing for our mission to deploy robots to factory floors, where industrial robot data is scarce and expensive to collect. We're now testing and deploying FLUX-mimic with manufacturing leaders like Audi USA, on complex, multi-step manipulation long considered impossible for conventional automation.show more

Read our blogpost:

@bfl_ai Let's go!

@bfl_ai 🦾🦾

@bfl_ai a very no-nonsense partnership that brings video models to life (and is euromaxxed)

@bfl_ai 🦾🚀🚀

@bfl_ai Cool stuff!

@_fracapuano @bfl_ai Now I’m really curious to know how they get accurate (enough) world dynamics representation with video models like FLUX (for robotics at least). JEPA makes a lot of sense to me but this makes me think twice. Any insights please @bfl_ai @elvisnavah ?

@bfl_ai 🤖🤖🤖

@bfl_ai wowowow

@chrisoffner3d @bfl_ai 🤩

@bfl_ai 🦾

@bfl_ai 🔥🔥🔥

Everyone is reacting to the dexterity but that is not the only shift. Fewer demos needed, running on one GPU and on premises is what changes who can train this. Factories without a teleop dataset just got unlocked. I'm curious about the 5% where the prediction is wrong and how fast do you catch it before the arm acts?"

@bfl_ai FLUX-mimic + robotics is insane 🤖 Turning video generation into real dexterity on a single GPU is the future

@bfl_ai lovely

@bfl_ai This is huge FLUX mimic real factory deployment? Robot learning just leveled up.

@bfl_ai 😍😍😍

@bfl_ai neat to see flux used for dexterity

@bfl_ai How come it's the other german car manufacturer that is exploring robotics outside of @BMW Definitely exciting tech tho!

@bfl_ai video-action models are still basically imitation learning with a video encoder slapped on top. not convinced that changes the core problem yet

@bfl_ai 👀

@bfl_ai Starting to test this in factories is great. Deployment means more data, and more data means a better model... It is a cycle. This deployment proves that responsive models can be used instead of just basic radio models.

@bfl_ai Lets goo!

Strong work. The core thesis is correct — control reduces to visual prediction, and sample efficiency on real hardware is the actual bottleneck. Factory floors with known parts and controlled lighting are still the easy mode. We’re running the harder version: dense outdoor polyculture, living deformable tissue, constant occlusion, shifting light, wind, and dirt. Same VAM principle, different physics. If FLUX-mimic can hold up under those conditions with the same data efficiency, that’s the unlock for field systems beyond manufacturing. The Reserve is built as the filter. Grit over gloss.

@bfl_ai Surely in the world of AI video generation, nobody is going to fund a startup based on a snazzy video, they are going to want to dive deep into the technical details, or is it a serious case of "I WAS BORN TO LEAD NOT TO READ"

@bfl_ai Training robots through video-action models feels like the practical path for factory work. The part I’m most curious about is how fast teams can adapt it to a totally new task.