Quickstart: FLUX 3 Action DROID

This quickstart serves the DROID checkpoint’s base BF16 package. It expects three RGB camera observations and an eight-value joint state. Review the checkpoint license before deployment; it is non-commercial.

PI0.5 YAM

Three camera images and a seven-value state.

Install and sign in

Create an environment and install servo-client:

your computer

python3 -m venv .venv
source .venv/bin/activate
python -m pip install servo-client

your computer

servo login

Open the login link in a browser. servo whoami confirms the active account.

Enroll the robot

On the robot computer, enroll it with the name robot-1:

robot computer

python3 -m venv .venv
source .venv/bin/activate
python -m pip install servo-client

Enroll it; the command prints a code and waits for approval:

robot computer

servo agent install --name robot-1

An organization admin approves the printed code on your computer:

your computer

servo robot approve CODE

The robot computer reports Enrolled robot robot-1. when it is ready.

Inspect the checkpoint

Inspect the pinned repository to confirm the input names and shapes:

your computer

servo model inspect hf://black-forest-labs/flux-3-action-droid --revision e3b6c915cdcdfb1d5db08e133b8ca3d29be2d7d4

The released DROID contract has three 640×360 RGB images (observation.images.wrist_image_left, observation.images.exterior_image_1_left, and observation.images.exterior_image_2_left) and an eight-value observation.state. Use the report for the tensor layout expected by your checkpoint.

Deploy the checkpoint

Deploy that same pinned checkpoint to the enrolled robot:

your computer

servo model deploy hf://black-forest-labs/flux-3-action-droid@e3b6c915cdcdfb1d5db08e133b8ca3d29be2d7d4 --robot robot-1 --wait

The default package uses the released BF16 weights and base sampling recipe. Servo reports the deployment ID and waits until it is ready.

Predict an action chunk

On the robot computer, load observations from your camera and robot code, then send them to the deployment:

predict_flux3.py

import numpy as np
import servo

sv = servo.Servo()
robot = sv.robots.get("robot-1")
deployment = sv.deployments.get("DEPLOYMENT_ID")
policy = deployment.policy(robot)
inputs = {
    "observation.images.wrist_image_left": np.load("wrist.npy"),
    "observation.images.exterior_image_1_left": np.load("left_exterior.npy"),
    "observation.images.exterior_image_2_left": np.load("right_exterior.npy"),
    "observation.state": np.load("joint_state.npy"),
}
with sv.session(policy, instruction="pick up the red cup") as session:
    chunk = session.predict(inputs=inputs)
print(chunk.horizon, "steps x", chunk.action_dim, "channels; nothing has moved")

The returned chunk contains 32 absolute joint-position actions. Prediction returns actions; your application remains responsible for safety checks and sending commands to the robot.

Stop the deployment

Stop it when you are done:

your computer

servo deployment stop --model flux-3-action-droid

Get started

Guides

Reference

Quickstart: FLUX 3 Action DROID

PI0.5 YAM