Fine Tuning

Extend Robotics supports model fine-tuning on our web console. Fine-tuning allows you to adapt an existing pre-trained model to a new dataset. This is useful when you want to transfer the model’s know

When to Use Fine-Tuning

• You have a new dataset that differs from the training data of the pre-trained model.

• You want to improve performance on domain-specific tasks (e.g., industrial object manipulation, agriculture robotics).

• You have limited data and prefer to leverage a strong base model rather than training a new one from zero.

Steps to Fine-Tune a Model

Step 1: Select model

Navigate to the Model Training page on the web console. Choose the pre-trained model you wish to fine-tune.

Step 2: Select Finetune Model to configure the parameters Here, you can:

• Review the base model’s current details

• Prepare parameters for fine-tuning

Step 3: Add new dataset

Step 4: Configure Hyperparameters

Optionally adjust fine-tuning parameters:

• Learning Rate — Lower values (e.g., 1e-5 to 5e-5) are safer for fine-tuning to avoid catastrophic forgetting

• Batch Size — Depends on dataset size and available GPU memory

• Epochs — Start with fewer epochs; monitor validation metrics

• Chunk Size — Match with motion sequence length (~1s recommended)

• KL Weight / Regularization — Adjust based on task precision requirements

You can also leave the default settings if unsure, as they are optimized for most use cases.

Step 5: Launch Fine-tuning

Click Create Training to launch the training job. You can monitor the training status in the training job dashboard.