Train and deploy built-in models
This topic describes how to train built-in VLA models on the platform and deploy trained models as inference services. This is the simplest approach. You do not need to create a development machine or customize a model. You can directly use built-in models to complete training and deployment.
Model library
The platform provides the following mainstream embodied intelligence models:
|
Model |
Description |
|
GR00T |
An NVIDIA general-purpose robot foundation model that supports multimodal understanding and task execution. |
|
PI 0.5 |
An embodied intelligence policy model designed for robot control and decision-making. |
In the left-side navigation pane, expand Models and click Model Library to browse all available models.
Train built-in models
-
In the left-side navigation pane, click Model Training.
-
Click Start Training.
-
In the dialog box that appears, confirm that Model Source is set to Official Model (selected by default).
-
Configure the following parameters:
Parameter
Required
Description
Dataset
Required
Select a dataset for training. You must register the dataset on the Datasets page in advance.
Base Model Type
Required
Select a model type from the drop-down list (
gr00torpi0).Select Model
Required
Select a specific model version.
Training Parameters
Required
Configure training parameters such as batch size, maximum training steps, and checkpoint save interval.
-
Click OK. The training task starts.
Monitor training
After a training task is submitted, you can view the task status and progress in the model training list:
|
Column |
Description |
|
Model Name |
The name of the model produced by training. |
|
Base Model Type |
GR00T or PI 0.5. |
|
Dataset Name |
The dataset used for training. |
|
Ray Dashboard |
Click the link to open the Ray Dashboard, where you can view detailed metrics of distributed training (including Loss curves, learning rates, etc.). |
|
Status |
Training, Completed, or Failed. |
To view training metrics (Loss curves, learning rates, etc.), go to More > View Metrics in the training task list. You must specify the TensorBoard log path when submitting the training task.
Deploy inference service
Trained models can be deployed as cloud-based inference services with one click.
-
In the left-side navigation pane, click Model Deployment.
-
In the model list, find a trained model with the Idle status.
-
Click Deploy Inference Service.
The deployment process runs automatically without additional configuration. The platform uses Ray Serve to deploy model inference services. You can view deployment logs, GPU resource usage, and service status through the corresponding Ray Dashboard.
After the deployment is complete:
-
The model status changes from Idle to Deployed.
-
The Inference Service URL column displays the access URL of the inference service.
Regardless of whether you deploy a GR00T or Pi0 model, deploying an inference service consumes 1 GPU resource.
To stop the inference service, click Stop Inference Service. The system stops the current inference service and releases the occupied GPU resources.
The statistics cards at the top of the page display the number of models in each status: Deployed, Deploying, Idle, and Other.
Edge-cloud collaborative inference
After the inference service is in the Deployed status, you can connect the cloud-based inference service to the robot runtime through the platform for edge-cloud collaborative inference. For robots that are compatible with LeRobot, you can quickly set up the edge-cloud inference pipeline. For other control protocols or data formats, you can perform secondary development based on the adb-robot-client SDK.
Procedure
-
Go to the robot deployment page: On the Model Deployment page, click a model with the Deployed status and select Robot Deployment.
-
Add a robot device: Click Add Device and select an online robot that is registered on the platform. The system automatically detects the device online status.
-
Initialize the device: Click Connect. The platform sends the inference service connection information and the first frame of the training dataset to the robot to initialize the robot state. After initialization, the inference status changes to Ready.
-
Start inference: Click Start Inference. The robot continuously sends requests to the inference service and begins edge-cloud collaborative inference.
-
Stop inference: Click Reset to stop the inference process and restore the initial state.
Edge-cloud collaborative inference requires an online device. For information about how to connect devices, see Manage devices and robot access.