Isaac05 quickstart

Use Isaac 0.5 with Servo

This guide deploys the public PerceptronAI/Isaac-0.5 checkpoint through Servo’s hosted model path. It assumes you have completed the Servo quickstart through robot enrollment, signed in with servo login, and connected your Hugging Face account. Servo has a checkpoint-native Isaac 0.5 runtime. You pass the Hugging Face repository to Servo; you do not need to clone the Perceptron inference repository or install its model dependencies on the robot computer. Servo pins the released checkpoint revision below so the inspected inputs and the deployed model refer to the same files. This is a large, high-memory policy, so deployment depends on suitable hosted capacity.

Inspect the checkpoint

Inspect the pinned checkpoint revision and read Servo’s reported input contract.

your computer

servo model inspect hf://PerceptronAI/Isaac-0.5 \
  --revision 6e397429f0ab91528ebac915004eba5e763fcdd4

The checkpoint’s input contract is two RGB camera images and an eight-value state vector. For this revision, Servo reports observation.images.image and observation.images.wrist_image as uint8 images shaped 256 × 256 × 3, and observation.state as float32 shaped [8]. Use the inspection output as the authority if you choose a different revision.

Import the checkpoint

Import the pinned checkpoint from Hugging Face into your Servo organization. This downloads and prepares the model; the command waits for it to become ready by default.

your computer

servo model import hf://PerceptronAI/Isaac-0.5 \
  --revision 6e397429f0ab91528ebac915004eba5e763fcdd4

Copy the imported model ID from the command output. You can find it again with servo model list.

Deploy near your robot

Deploy the imported model near the enrolled robot. Isaac 0.5 is a large, high-memory policy, so deployment depends on suitable hosted capacity and its initial startup takes longer than a small checkpoint.

your computer

servo model deploy mdl_YOUR_IMPORTED_MODEL_ID --robot robot-1

Save the deployment ID printed by Servo. Wait for the deployment to report active before sending predictions. Replace robot-1 with the name or ID of your enrolled robot.

Send the first observation

Use real sensor arrays with the names and shapes from inspection; do not send test data to a physical robot.

isaac_first_prediction.py

import numpy as np
import servo

sv = servo.Servo()
deployment = sv.deployments.get("dep_YOUR_DEPLOYMENT_ID")

inputs = {
    "observation.images.image": np.load("image.npy"),
    "observation.images.wrist_image": np.load("wrist_image.npy"),
    "observation.state": np.load("state.npy"),
}

with deployment.session(
    instruction="pick up the object and place it at the target",
) as session:
    prediction = session.predict(inputs=inputs)
    print(f"Actions: {len(prediction.actions)} steps x {prediction.action_dim} values")

The arrays must match the names, shapes, and dtypes reported by servo model inspect. Connect the returned action chunk to a robot-side controller that understands Isaac’s action convention and control rate; Servo does not infer a robot-specific calibration from the checkpoint.

Stop the deployment

Stop the deployment after use:

your computer

servo deployment stop dep_YOUR_DEPLOYMENT_ID --wait

The imported model remains in your organization for later deployments. For model details and the checkpoint’s license and reference implementation, see the Hugging Face model card.

Get started

Guides

Reference

Isaac05 quickstart

Use Isaac 0.5 with Servo

Inspect the checkpoint

Import the checkpoint

Deploy near your robot

Send the first observation

Stop the deployment

​Use Isaac 0.5 with Servo

​Inspect the checkpoint

​Import the checkpoint

​Deploy near your robot

​Send the first observation

​Stop the deployment

Use Isaac 0.5 with Servo

Inspect the checkpoint

Import the checkpoint

Deploy near your robot

Send the first observation

Stop the deployment