Skip to main content

​
Overview

SmolVLA is a lightweight Vision-Language-Action model optimized for edge deployment. Cyberwave supports fine-tuning SmolVLA models on your custom datasets using the LeRobot v3 dataset format.
SmolVLA models use the LeRobot v3 dataset format for training, which differs from the TFDS format used by OpenVLA models. The platform handles format conversion automatically.

​
Model Selection

When starting a new training:
  1. Navigate to AI → Training in your environment
  2. Select SmolVLA from the available model architectures
  3. The platform will automatically use the LeRobot training pipeline
Only models with is_trainable: true appear in the training model selection. SmolVLA is pre-configured as trainable.

​
Dataset Conversion

When you start training with a SmolVLA model, the platform:
  1. Joins your episode parquet files into a single dataset
  2. Converts the OpenVLA-format parquet to LeRobot v3 format
  3. Handles camera role mapping (primary, wrist, secondary)
  4. Encodes video frames using AV1 codec (configurable)
The conversion is cached — subsequent trainings with the same dataset and configuration skip the conversion step.

​
Training Parameters

SmolVLA training supports:

​
Camera Configuration

Camera roles are mapped to LeRobot conventions:
  • primary → observation.images.primary
  • wrist → observation.images.wrist
  • secondary → observation.images.secondary
Configure camera roles in the training wizard or via the API.

​
Training Workflow

​
Deployment

After training completes:
  1. Deploy the trained model as a controller policy
  2. Assign the VLA controller to your robot twin
  3. Use natural language prompts to control the robot
SmolVLA models are optimized for edge inference, making them suitable for real-time robot control with lower latency than larger VLA models.

​
Reuse camera setup

Open Camera setup on a model to review its cameras and mounting positions in another environment. Saved asset and twin slugs help match resources across installations. Resolve missing cameras or changed assets before applying, then restart simulation. Model owners can also save an environment setup as a new recommendation while retaining the setup recorded with training data. The same training flow has experimental support for π0 (LeRobot) and π0.5 (LeRobot). Fine-tune for your robot and review its camera and joint requirements; a base policy is not a ready-to-run SO101 controller. Full π-model training and task success have not yet been validated in this integration.