Image Lab Tops Robot Leaderboard: FLUX 3 Action Runs on a Desk-Size Arm
lmoroney · x · 2026-09-30
Black Forest Labs, the team behind the FLUX image models, released FLUX 3 Action, an open-weights 7B world action model: it takes a camera frame plus a text instruction and returns the next 2 seconds of actions.
- Architecture: a diffusion transformer that denoises video tokens and action tokens together, with a frozen video VAE and a frozen Qwen3-VL-4B instruction encoder; each call returns 32 actions and optionally 32 predicted frames.
- Control loop: the model imagines future frames while deciding how to move; at control time you skip frames, execute some actions, look again, and replan.
- Checkpoints: base model, a DROID fine-tune for the Franka arm, and an SO-101 fine-tune; fine-tunes on two small video games and an indoor drone trained in NVIDIA Isaac Sim suggest it's meant as a general policy you adapt to your own tasks.
- Leaderboard: first place on NVIDIA's RoboLab-120 at 42.9% (515/1200 trials), ahead of HiDream-O1-Embodied (39.9%), Atomic-WAM (39.6%), and OASIS (39.0%).
The author also covers caveats worth knowing before deploying it, and notes you can LoRA it onto a desk-size arm with LeRobot.
More from Embodied
- Astribot releases Lumo-2 tech report: three-stage aligned training, label-free skill transfer — CyberRobooo · 2026-09-30
- Astribot demos two humanoid robots collaborating with its Lumo-2 world-action model — CyberRobooo · 2026-09-30
- Fully open-source VLA MolmoAct2 lands CoRL 2026 Spotlight, matching π0.5 performance — jmin__cho · 2026-09-30
- Apple's HomePad smart home hub reportedly launches October 13th alongside new HomePod mini and Apple TV — tomwarren · 2026-09-30
- Open-source self-driving guide adds VFH: binning LiDAR into 5° obstacle histograms — 4310sy · 2026-09-30
- FreeCAD AI Bridge Lets Any AI Model Build 3D Models in FreeCAD — poofph · 2026-09-30