isaac_ros_graspgenx_models_install#

Source code available on GitHub.

Model setup for the isaac_ros_graspgenx package.

install_graspgenx_models.sh reads the GraspGenX ONNX model, its external weights, and model_manifest.json from ${ISAAC_ROS_WS}/isaac_ros_assets/isaac_ros_graspgenx/model. The script:

  1. Verifies each model file against its SHA-256 checksum in model_manifest.json.

  2. Builds the native TensorRT plugin, libgraspgenx_plugins.so, for the target GPU.

  3. Builds the FP16 TensorRT engine with trtexec.

The script writes these files to ${ISAAC_ROS_WS}/isaac_ros_assets/models/graspgenx:

File

Content

graspgenx_everything_fp16.engine

TensorRT engine for the target GPU

libgraspgenx_plugins.so

TensorRT plugin that the engine requires at runtime

engine_build.json

Build signature of the model, plugin, script, trtexec, and GPU

Run the installer with --eula to review and accept the model license before building the TensorRT engine.

The script reuses the installed engine when the build signature is unchanged. Follow the sequence in Using Hugging Face checkpoints below.

Using Hugging Face checkpoints#

Download GraspGenXModel checkpoints and export ONNX using the model code bundled in this package’s vendor/ directory. The exporter automatically downloads both generator and discriminator checkpoints and their configurations.

export HF_GGX="${ISAAC_ROS_WS}/isaac_ros_assets/isaac_ros_graspgenx"

The export writes the ONNX model, external weights, and model_manifest.json to ${HF_GGX}/model, the installer’s default input directory. Keep ISAAC_ROS_WS unchanged between export and model setup.

Install the export dependencies and run the exporter for your installation method:

Prerequisite: complete package installation and ROS environment sourcing in the Binary Package tab of Build isaac_ros_graspgenx. If you arrived here from the quickstart, these steps are already complete.

  1. Prepare the export dependencies once per export environment. Skip this step if ${HF_GGX}/venv is already prepared:

    ros2 run isaac_ros_graspgenx_models_install install_graspgenx_models.sh --install-export-deps
    

    This command only prepares the export environment; it does not build the model or engine.

  2. Download both checkpoints and export them to ONNX:

    ros2 run isaac_ros_graspgenx_models_install export_huggingface_model.sh \
       --checkpoint-dir "${HF_GGX}/checkpoints" \
       --output-dir "${HF_GGX}/model"
    
  3. After the export finishes successfully, review and accept the model EULA when prompted, then build the TensorRT engine and plugin:

    ros2 run isaac_ros_graspgenx_models_install install_graspgenx_models.sh --eula
    

Continue with the runtime asset paths in the quickstart. Complete the export before building the model installation package from source.