Black Forest Labs co-founder Andreas Blattmann says the company has built a 7-billion-parameter model that predicts robot movements alongside video, with factory-floor use in mind. The specific robotics work is a striking turn from image generation, and comes with a practical claim: one industrial task reportedly needed about four hours of fine-tuning data.
Black Forest Labs Watch analysis
What happened
In an interview with the Latent Space podcast, Blattmann and fellow co-founder Dustin Podell discussed the company’s research and plans. BigGo Finance’s account says the model predicts robot joint positions and rotations from recent camera frames and an instruction, then updates its predictions as new footage arrives. It runs at roughly 10 frames per second and predicts short chunks of around nine to 21 frames.
Blattmann says one industrial application reached an acceptable success rate with about four hours of data, roughly eight times less than the open-source baseline he identified as pi 0.5. The account does not provide the task, test conditions or a fuller comparison, so that is an attributed claim rather than a general benchmark result. The founders also expect initial commercial robotics applications in manufacturing in 2027.
Why it matters
A relatively small model intended to run in real time on a factory floor points towards a different route to robotics than sending every decision to a large cloud system. The reported fine-tuning requirement, if borne out across tasks, could also make adapting robots to specific work less data-hungry.
This is a development in robotics, not just a company’s change of direction in an interview. But the technical description has limits: the system currently has no high-level planner in the loop, and the founders acknowledge that closing the gap between simulation and real-world performance remains work in progress.
Our read
The interesting news is not that Black Forest Labs wants to build robots. It is the concrete recipe: a 7B model predicting motion with video, on-device ambitions and a reported four-hour fine-tuning result. That is enough to take seriously, not enough to declare the factory floor solved. The next evidence should come from real deployments, not another prediction with a calendar attached.
What to watch
- Whether Black Forest Labs announces a manufacturing deployment or commercial robotics product in 2027.
- Independent details on the task, conditions and baseline behind the four-hour fine-tuning claim.
- How the model handles planning and the simulation-to-real-world gap.
Discussion spark: Should factory robots prioritise compact models that can run on-device, even if they need a separate planning system, or is one integrated model the better route?
Sources and evidence
- Black Forest Labs' Blattmann: Image Generation Was a Detour, Robotics Is the Destination – BigGo Finance (17 November 2026, 11:09 UTC)
not affiliated with or endorsed by Black Forest Labs