Skip to content

Hardware Workflows

Hardware-specific workflows are tailored for users who have an EdgeFirst-supported target device and want to go through the full MLOps loop: device setup, dataset acquisition, model training, model conversion, on-target validation, and deployment.

Select your platform from the table below to get started.

About these workflows

The hardware workflows use the Coffee Cup sample dataset and train a ModelPack vision model. Cost estimations shown in the table reflect this specific configuration — actual costs will vary depending on your dataset size, the model type, and training configurations.

Persona Hardware Features Cost Time
Maivin PC + Maivin Record MCAP, Annotate 2D, Train, Convert to TFLite, Validate on Target, Deploy on Maivin TBA TBA
Raivin PC + Raivin w/ Radar Record MCAP, Annotate 2D + 3D, Train, Convert to TFLite, Validate on Target, Deploy on Raivin TBA TBA
LiDAR (coming soon) PC + Raivin w/ LiDAR Record MCAP, Annotate 2D + 3D (enhanced), Train, Convert to TFLite, Validate on Target, Deploy on Raivin TBA TBA
i.MX 8M Plus PC + i.MX 8M Plus Copy Dataset, Train, Convert to TFLite or Kinara (FRDM), Validate on Target, Deploy on i.MX 8M Plus ~$9-11 USD ~1.5-3.2 hours
i.MX 95 PC + i.MX 95 Copy Dataset, Train, Convert to Neutron or Kinara (FRDM), Validate on Target, Deploy on i.MX 95 ~$8-9 USD ~1-1.2 hours
Jetson Orin PC + Jetson Orin Copy Dataset, Train, Convert to TensorRT, Validate on Target, Deploy on Jetson Orin ~$8 USD ~1 hour
Raspberry Pi 5 PC + Raspberry Pi 5 Copy Dataset, Train, Convert to Hailo, Validate on Target, Deploy on Raspberry Pi 5 ~$9-12 USD ~1.8-4.3 hours

  1. Maivin Workflow

    This workflow explores recording MCAPs from the Maivin to create and annotate datasets with 2D bounding boxes and segmentation masks. Once annotated, you will train and validate a Vision model, then deploy it back to the Maivin for inference.

  2. Raivin Workflow

    This workflow explores recording MCAPs from the Raivin to create and annotate datasets with 2D and 3D annotations. Once annotated, you will train and validate a Fusion model, then deploy it back to the Raivin for inference.

  3. LiDAR Workflow (coming soon)

    An extension of the Raivin workflow for users with a LiDAR sensor integrated on the Raivin. Adds enhanced 3D annotation and Fusion model capabilities.

  4. i.MX 8M Plus Workflow

    This workflow explores copying the Coffee Cup dataset from "Sample Project" and training a Vision model that detects coffee cups. Once trained, you will convert the model for the i.MX 8M Plus NPU, validate on target, and deploy it.

  5. i.MX 95 Workflow

    This workflow will cover copying a dataset, training a Vision model, converting it for the i.MX 95 eIQ Neutron NPU, and deploying it on device.

  6. Jetson Orin Workflow

    This workflow explores copying the Coffee Cup dataset from "Sample Project" and training a Vision model that detects coffee cups. Once trained, you will convert the model to TensorRT, validate it, and deploy it on the Jetson Orin.

  7. Raspberry Pi 5 Workflow

    This workflow will cover copying a dataset, training a Vision model, converting it for the Raspberry Pi 5 with Hailo-8L NPU, and deploying it on device.

Future Work

The workflows with missing links are a work in progress and currently unavailable.