FLUX 3 Action DROID
FLUX 3 Action DROID is a 7B policy adapted to DROID-style robot observations. It uses three cameras and robot state to generate joint commands and future visual frames.
Model details
The input uses wrist, left and right cameras. State contains seven arm joint positions in radians and a gripper closed fraction. Outputs are absolute joint targets plus gripper state, with 32 actions at 15 Hz. The repository provides separate guidance- and step-distilled variants.FLUX 3 Action DROID official documentation
Practical use
Practical guidance. Keep DROID and SO-101 integrations separate because their camera layouts and action conventions differ. Pin the shared encoders with the policy revision. Review trajectory limits and controller behavior independently of predicted image quality before moving from simulation to hardware.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | black-forest-labs/flux-3-action-droid |
| Weight license | FLUX Kommunity License 1.0; text encoder separately under Apache 2.0.FLUX 3 Action DROID official documentation |
| Default package | BF16 policy with saved processors.FLUX 3 Action DROID official documentation |
| Variants | Four-step base and guidance-distilled packages; one-step distilled package.FLUX 3 Action DROID official documentation |
Related models
- FLUX 3 Action Base
Black Forest Labs · FLUX 3
Seven-billion-parameter world-action model that jointly predicts future visual frames and robot action sequences.
Launch date not confirmedDownloadable Seven-billion-parameter SO-101 robot policy that predicts actions and future frames from cameras, robot state and task instructions.
Launch date not confirmedDownloadable