Flux 3 Action: A 7B open-weight world action model for robots
7 hours ago
- FLUX 3 Action is an open weights world action model (WAM) that jointly predicts future frames and actions.
- It achieved a 42.92% success rate on the RoboLab benchmark, surpassing the 16B Cosmos 3 Nano Policy (36.8%).
- The model is a 7B parameter diffusion transformer trained on action data from multiple robot embodiments.
- Input includes camera frames, a state vector, and a caption; output consists of 32 actions and optionally 32 decoded frames.
- It is integrated with LeRobot for real robot arm control and offers parameter-efficient fine-tuning recipes.
- Fine-tuned for video games (GRUNT, VECTOR) and indoor drone control, showing generalization across tasks.
- Weights, code, and documentation are publicly available.