Skip to main content
FLUX 3 Action is available under its non-commercial license. The public checkpoint is trained for DROID data, but Servo deploys the policy to the robot you name with --robot; DROID is not a deployment requirement.

​
Checkpoint format for fine-tunes

Publish a complete FLUX 3 Action policy package in a Hugging Face repository. The reference package has root config.json, config.native.json, manifest.json, and model.safetensors; include the weight files selected by the manifest. The root config uses the LeRobot-style type: "flux3". Keep the checkpoint’s camera names and order, state and action widths, image layout, chunk size, control rate, normalization, and sampler recipe consistent with the fine-tuned weights. The importer reads this recipe from config.json; it does not infer or repair missing values. The checkpoint recipe includes camera_keys, camera_layout, canvas_hw, chunk_size, fps, action_scale, gripper_flip_dims, sampler, num_inference_steps, guidance values, sampler_shift, dtype, and the declared video VAE and text encoder IDs. Keep referenced encoder assets available at those IDs. The policy package’s config.json and policy_preprocessor.json must agree on the camera inputs. For the public checkpoint, the source revision is pinned below; for a fine-tune, inspect and deploy the exact Hugging Face commit you published.

​
Inspect the input contract

Inspection reads metadata and small config files, not model weights. The public checkpoint declares three RGB image arrays and an eight-value state array. Its image keys are observation.images.wrist_image_left, observation.images.exterior_image_1_left, and observation.images.exterior_image_2_left; inspection shows their shapes and the state key. A fine-tuned package can declare different inputs, so inspect that revision before connecting the robot’s cameras:
inspect_checkpoint.py

​
Deploy

Deploy the pinned public checkpoint to any enrolled robot supported by the policy contract:
your computer

​
Import and deploy a fine-tune

Import the fine-tune at the same pinned revision you inspected, then deploy the returned model ID:
your computer
The public checkpoint can be deployed directly from its hf:// path as shown above. For a fine-tuned package, keep its policy inputs, recipe, and referenced encoder assets together at the revision you import.

​
Send observations and receive actions

Use the contract’s exact keys and tensor shapes when constructing Session.predict(inputs=...). The quickstart prints the deployed policy’s resolved keys, shapes, and dtypes before its prediction example.