Choose a model

Point Servo at your checkpoint, read what it expects from your robot, and reach a private bucket.

Servo runs the model you already have. This page covers the handle you get from a path, what inspection tells you before anything is deployed, and how to reach a private bucket.

Point Servo at the checkpoint

Give Servo the path where the checkpoint lives:

your computer

import servo

sv = servo.Servo()
model = sv.models.get("hf://allenai/MolmoAct2-BimanualYAM")

get returns at once and pins the current content of that path; the weights are fetched when you deploy. Use s3:// or gs:// for a checkpoint in your own bucket, and revision= to pin a Hugging Face commit or branch. A handle never changes under you, so call get again to pick up a newer revision. sv.models.list() returns every model your team has pointed Servo at with get.

Inspect it against your robot

Resolve the model’s contract for one registered robot:

your computer

robot = sv.robots.get("yam-cell-01")
report = model.inspect(robot)
print(report.compatible, report.reasons)

Inspection reads the checkpoint’s metadata and the robot’s registered type; it runs no inference, allocates nothing, and needs no connected hardware. compatible is false when the robot type or its camera and state layout does not match, and reasons says why. A compatible model still predicts nothing until deploy runs it and start binds it to a robot, as the quickstart shows.

What the model expects

The report describes the inputs your robot supplies and the chunk that comes back, here for the YAM checkpoint:

Field	For MolmoAct2 Bimanual YAM
cameras	top, left, and right, each with the size the checkpoint was trained at; send frames as captured and the model’s own processor resizes them
state_dim	Fourteen values in training order: left arm joints, left gripper, right arm joints, right gripper
horizon, action_dim	Thirty rows of fourteen channels per chunk, as absolute joint positions
dt_s	The spacing between rows, which is the model’s cadence and not a request latency
channels	The name and unit of every action channel, in order
missing	Values the checkpoint’s metadata leaves unspecified, with the reason

Arm joints use radians. The checkpoint’s metadata leaves the joint zero and sign conventions and the physical gripper units unspecified, so confirm them with the model maintainer for a new robot. Send frames as captured and let the model’s processor resize them, because pre-resizing changes the input path. The upstream checkpoint metadata holds the normalization statistics.

What a complete export contains

Keep every file your exporter produced together: the weights, the config that declares the architecture and its features, and the processor files that define observation normalization and action unnormalization. Upload the directory, or an archive of it, to your bucket. If inspection reports an unsupported model family, contact the Servo team with the path.

Private buckets

For a public path, skip this section. For a private bucket or a gated Hugging Face repository, register the source once from your computer, and get resolves it from the path prefix.

Hugging Face

Make a read token available as HF_TOKEN in your environment, then register and verify the repository:

your computer

servo source add hf://allenai/MolmoAct2-BimanualYAM --token-env HF_TOKEN
servo source verify hf://allenai/MolmoAct2-BimanualYAM

Amazon S3

Register the AWS role that can read the bucket, give it the trust relationship the command prints, then verify:

your computer

servo source add s3://lab-checkpoints/ --aws-role arn:aws:iam::123456789012:role/servo-reader
servo source verify s3://lab-checkpoints/

Google Cloud Storage

Register the Google Cloud identity that can read the bucket, complete the workload identity trust the command prints, then verify:

your computer

servo source add gs://lab-checkpoints/ --service-account [e-mail]
servo source verify gs://lab-checkpoints/

Verification checks the credential and the trust; get checks access to the files at the path. No secret ever appears in a path.

You now have

• A model handle pinned to one revision.
• A compatibility report for your robot, read before any capacity exists.

Next

First motion

Preflight without motion, then a guarded 30-second episode.

Keep your own loop

The two calls under run, and a session that keeps its connection.