Model Training

Extend Robotics provides model training on the web console. This guide will walk you through how to train a model on your dataset, monitor your training and deploy the model.

Getting Started

To begin, navigate to the Model Training section from the left-hand menu on the web console. All of your existing training jobs will be listed here.

See the Video Guide for training creation.

Creating New Training Job

Step1: Enter Training ID and Select Dataset

To create a new training job:

1. Click on the New Training button in the Command Console Model Training page.

1. A pop-up window will appear. Here, you'll need to:

• Enter a unique name for your training job.

• Select one or more datasets.

• Choose a model architecture.

Tip: You can select multiple datasets to train your model, but make sure the selected datasets are from collected from the same setup.

Step 2: Select and Configure Model

We currently support the following model types:

• Action-Chunking with Transformer (ACT)

• ConvNet + Multilayer Perceptron (CNN-MLP)

Each model has different parameters and use cases. For a detailed explanation of model architecture and training configuration, refer to the full model documentation

Step 3: Start Training

Once you have configured your model, you are ready to start your training job. By clicking on Create Training, your training job will be submitted for training.

You can always check the status of your training job by navigating to the training job you created in Model Training page:

• In the Overview tab, you’ll see the current training status.

• In the Metrics tab, you can track loss curves, including components like KL divergence.

The training job status refreshes every 30 seconds. You can also click the Refresh button to update manually.

Resume Training

If a training job was stopped before completion, you can easily resume it without starting from scratch. When resuming a job, you can update certain hyperparameters to refine the process. Click Resume Training in the Training Job dashboard, then you can update some of the hyperparameters in the panel.

Train More

After a training job is completed, you may decide to extend training if you believe the model can benefit from additional epochs. This allows you to continue improving performance without starting a new job from scratch.

Navigate to the Training Jobs dashboard, click Train More, and configure the Number of Epochs you would like to add. The model will resume training from the last saved checkpoint and continue for the newly specified epochs.