DownloadFLUX.3 Action SO-101 is a 7B open-weight world action model from Black Forest Labs, specialized for controlling SO-101 robotic arms. It combines visual observations, robot state information, and natural-language instructions to predict upcoming robot actions while simultaneously denoising predicted future video frames. The checkpoint uses two camera streams, scene and wrist, alongside six-dimensional state and action representations and historical observations for temporal context. It predicts chunks of 42 actions, executes 32 at 30 Hz, and then replans, enabling closed-loop robotic control. The model was trained on SO-101 episodes from the LeRobot Community Dataset v3 and integrates directly with the LeRobot framework. Developers can adapt it to new robotic tasks using the provided rank-32 LoRA training recipe. The checkpoint uses BF16 weights and includes its own observation history, camera configuration, preprocessing pipelines, normalization statistics, and action contract.
Features
- 7B-parameter open-weight world action model
- Scene and wrist camera visual inputs
- Natural-language instruction following for robotic tasks
- Joint-delta actions with absolute gripper control
- 42-action prediction chunks with 32 actions executed at 30 Hz
- Rank-32 LoRA recipe for adaptation to new tasks
- Native integration with the LeRobot robotics framework
- Joint action and future video-frame prediction
Project Samples
