Black Forest Labs publishes FLUX 3 Action weights, a 7B world action model for robots
In one sentence On September 23 Black Forest Labs released weights, code and technical report for FLUX 3 Action, a 7B model that takes images, robot state and a text instruction and predicts the next 32 actions plus the scene's video. On RoboLab-120 BFL claims 42.92%, a figure that does not match others in the same report.
In July Black Forest Labs, the German lab known for the FLUX image generators, showed an early-access variant of its model aimed at robots. On September 23 it made the weights public: FLUX 3 Action is a 7-billion-parameter "world action model".
The model looks at camera images, reads the robot's state and a written instruction, and produces two things: the arm's next 32 moves and a prediction of how the scene will change. The idea is that a model trained mostly on video already has an intuition of real-world physics that can be reused to move real objects.
The base model is not a ready-made robot: it has to be adapted with teleoperation data from your own arm. In a real-robot test run with Positronic Robotics on a Franka arm it succeeded in 28 of 30 attempts. According to gradually.ai the weights are under the FLUX Kommunity License, not a standard commercial license.
Companies
Black Forest Labs
Tools
FLUX 3 Action, LeRobot
Tags
Sources