FLUX 3 Action Tops RoboLab-120 at 42.9%, Bringing Video Pretraining to Robotics
DeepLearningAI · x · 2026-10-15
Black Forest Labs released FLUX 3 Action, a 7B-parameter world-action model that turns camera feeds and text instructions into robot movements, also predicting future camera views — unlike VLAs such as π0.5 and GR00T that predict actions only.
Key facts:
- Ranks first on Nvidia's RoboLab-120 simulation benchmark with a 42.9% success rate across 1,200 trials
- Pretrained on >95% video tokens, with images next and <0.5% audio
- Adapts to a new robot arm with about 200 demonstrations; Franka and SO-101 fine-tunes available
- Base and fine-tuned models are open on Hugging Face under the FLUX Kommunity License (commercial use limited to companies under $5M annual revenue)
- None of the benchmark leaders, including this model, have yet been tested on real robots via the independent RoboArena
More from Embodied
- Words to World: A Benchmark Testing LLMs on Closed-Loop Real-World Physical Control — Xinnan Dai · 2026-10-15
- Researcher muses on building a RoboParty robot from scratch in his own machine shop — andrew_n_carr · 2026-10-15
- Zipline shares stunning sunrise shot of two delivery drones at test site — AnandSwa · 2026-10-15
- Reflex Robotics demos snow-shoveling robot; blower founder jokes he's pivoting — MatthewChang · 2026-10-15
- Bagel Labs demo shows small WorldDiT model running a robot making coffee autonomously — aryanXmahajan · 2026-10-15
- One prompt gets Argon 4 to build a full Three.js kinematics sim for the ODRI Solo robot leg — Tight-Instruction-17 · 2026-10-15