Skip to content

Running Autoware without CUDA#

Although CUDA installation is recommended to achieve better performance for object detection and traffic light recognition in Autoware Universe, it is possible to run these algorithms without CUDA. The following subsections briefly explain how to run each algorithm in such an environment.

To install Autoware without the NVIDIA libraries (CUDA, cuDNN, and TensorRT), run the Ansible playbook with --skip-tags nvidia. See the source installation page. The packages that need CUDA or TensorRT check for these libraries at build time. If CMake does not find them, the packages leave out their GPU code.

The launch files on this page are in the autoware_launch package of the autoware_launch repository (version 0.53.0 in Autoware 2.0.0).

Running 3D object detection without CUDA#

The lidar_detection_model argument of component_perception.launch.xml selects the LiDAR object detector. Its value has the form model_type/model_name. If you leave out the model_name part, the default parameters of the model are used. The model types are:

  • centerpoint (the default)
  • transfusion
  • bevfusion
  • apollo
  • clustering

The description of the argument also names pointpainting. The pointpainting detector needs camera images, and it runs only in the camera-LiDAR pipeline. component_perception.launch.xml does not start this pipeline. As a result, pointpainting starts no neural network detector here.

All model types except clustering run a neural network with TensorRT and need CUDA. With clustering, Autoware starts no neural network detector. It uses only the rule-based detector, which is built on the euclidean_cluster package and runs on the CPU.

The top-level launch files, for example autoware.launch.xml and logging_simulator.launch.xml, do not pass lidar_detection_model to the component launch file. To select clustering, change the default value in autoware_launch/launch/components/component_perception.launch.xml:

<arg name="lidar_detection_model" default="clustering" description="If the model name is not declared, the default model in the detection composition will be used"/>

Keep the other GPU options of the same file at their default value false:

  • use_cuda_ground_segmentation: the CUDA backend of ground segmentation
  • use_semantic_segmentation_ptv3: the PTv3 neural network for ground segmentation
  • cuda_pointcloud_preprocessing: the CUDA backend of the common pointcloud downsampling

If you did not build the workspace with --symlink-install, rebuild autoware_launch after you change a launch file.

Running traffic light detection without CUDA#

Traffic light recognition has two steps: detection finds the traffic lights in the camera image, and classification finds their color and arrows. The launch file component_traffic_light.launch.xml starts both steps.

To run traffic light detection without CUDA, set the traffic_light_recognition/use_high_accuracy_detection argument of autoware.launch.xml to false. Then the YOLOX-based detectors (autoware_traffic_light_fine_detector or the whole-image detector), which need CUDA, do not start. Instead, the autoware_traffic_light_map_based_detector package calculates where the traffic lights appear in the image from the vector map. The default value of this argument is true in autoware.launch.xml. e2e_simulator.launch.xml always sets it to false. For other launch files, change the default value in autoware.launch.xml:

<arg name="traffic_light_recognition/use_high_accuracy_detection" default="false" description="enable to use high accuracy detection for traffic light recognition"/>

To run traffic light classification without CUDA, set the classifier type to 0 (HSV filter) in component_traffic_light.launch.xml. There is one argument for car traffic lights and one for pedestrian traffic lights:

<arg name="traffic_light_recognition/classification/car/classifier_type" default="0" description="0=HSVFilter, 1=CNN, 2=LAMPClassifier"/>
<arg name="traffic_light_recognition/classification/pedestrian/classifier_type" default="0" description="0=HSVFilter, 1=CNN, 2=LampRecognizer"/>

The default value 1 selects a CNN classifier, and 2 selects a lamp recognizer. Both use TensorRT and need CUDA. The HSV filter classifies the color of the light in the HSV color space and runs on the CPU. For details, see the autoware_traffic_light_classifier README.