Skip to content

black-forest-labs/flux-3-action-droid

Embedl's build embedl/flux-3-action-droid of the original model black-forest-labs/flux-3-action-droid

embedl/flux-3-action-droid architecture graph
embedl/flux-3-action-droid architecture graph
  • embedl/flux-3-action-droid
  • FP8
  • H200
Labels
  • Text
  • VLA
  • Robot Policy
Build
Embedl-optimized
Precision
FP8
Runtime
TensorRT 10.16.1.11 with Embedl optimizations
Action chunk
32 joint-position actions per call, 1 denoising step
Checkpoint
black-forest-labs/flux-3-action-droid, variants/sd (step-distilled)
Run FLUX 3 Action DROID on H200

FLUX 3 Action is Black Forest Labs' open-weights 7B world action model: it takes camera frames, the robot's state and a text instruction, and returns the next chunk of actions. This is its step-distilled DROID checkpoint, driving a Franka arm in the RoboLab simulator.

Simulator: RoboLab

ad45d4f · Isaac Sim 5.0, Isaac Lab 2.2.0 · Franka Emika Panda (DROID) · 15 steps per second

7 scenes

  • Banana in the bowl banana_bowl

    RoboLab-120 · at most 750 steps

    • Pick up the banana and place it in the bowl
    • Put the fruit in the bowl
    • Grasp the yellow banana and place it inside the red bowl on the table
  • An orange into the white bowl orange_bowl

    RoboLab-120 · at most 900 steps

    • Move an orange to the white bowl
    • Move orange to white bowl
    • Pick up one orange citrus fruit and place it inside the white bowl
  • Yellow block on red yellow_on_red

    RoboLab-120 · at most 900 steps

    • Stack the yellow block on the red block
    • Stack blocks: yellow on red
    • Pick up the yellow block on the left and place it precisely on top of the red block in front of the birdhouse
  • Red things into the bin red_items_bin

    RoboLab-120 · at most 900 steps

    • Put all the red things in the grey bin
    • Sort items by color
    • Identify every red-colored object on the table and place each one into the grey bin
  • Blocks into the bin blocks_in_bin

    RoboLab-120 · at most 2250 steps

    • Sort all colored blocks into the bin
    • Put away the colored blocks
    • Put away the solidly-colored wooden blocks from the table and place them all into the bin
  • Block tower block_tower

    RoboLab-120 · at most 1350 steps

    • Stack the blocks into a tower
    • Build a tower
    • Pick up the colored blocks and stack them vertically on top of each other, forming one tower
  • Blocks in color order colored_blocks

    RoboLab-120 · at most 1350 steps

    • Stack the blocks in the order from bottom to top: red, blue, green, yellow
    • Stack in the order of red, blue, green, yellow
    • Build a tower by placing the red block first, then the blue block on top, then the green, and finally the yellow block on top as a single tower

Not for commercial use: the weights, and this engine, are under the FLUX Kommunity License v1.0 (bfl.ai/licensing for other uses); flux-action code is Apache-2.0. The model outputs joint targets with no bounds on velocity, force or workspace: enforce those in the application. Not an official Black Forest Labs product.

Methodology

  • Latency: one DiT step of the step-distilled policy in trtexec with CUDA graphs; the text encoder, video VAE and sampler around it run in PyTorch and are not included.
  • Accuracy: task success is RoboLab-120's as Black Forest Labs publishes it for the step-distilled checkpoint, not measured on this build. The build checks one DiT step of the engine against PyTorch on a public DROID observation (action output, max abs).
  • Simulator: RoboLab (NVIDIA Isaac Sim) runs on its own device; the policy's runner drives it one episode at a time, so a run may wait for another.

Files

Image Size

Model

docker.io/embedl/models:hf.black-forest-labs.flux-3-action-droid.fp8-modelopt.h200.2026101814036bb4

—

Runtime

docker.io/embedl/models:flux-3-action.h200.runtime.amd64.20261021073351ba

—