Run autonomous driving simulations with CARLA on Deadline Cloud

The autonomous_driving_carla job bundle runs a CARLA autonomous driving simulation parameter sweep on AWS Deadline Cloud with configurable multi-sensor capture. The job runs a lane-change cut-in scenario and sweeps across configurable ego speeds, NPC speeds, and NPC starting distances. Each parameter combination creates a task (default 2×2×2 = 8 tasks).

Each task captures multi-sensor data from user-selected camera viewpoints and produces per-camera videos, RGB and semantic segmentation frames, LiDAR point clouds (.ply), and 2D/3D bounding boxes in KITTI format.

The job runs inside a Docker container based on carlasim/carla:0.9.16. Before submitting, build the custom Docker image from the docker/ directory in the bundle and push it to Amazon Elastic Container Registry (Amazon ECR). Attach the docker_nvidia_container_toolkit host configuration script to your fleet to provide Docker and the NVIDIA Container Toolkit.

To run this bundle, you need a Deadline Cloud farm with a GPU fleet (minimum 1 NVIDIA GPU, 16 vCPU, 64 GiB memory) and a conda queue environment with ffmpeg. The queue role needs Amazon ECR pull permissions.

From the job_bundles directory, submit the job:

deadline bundle gui-submit autonomous_driving_carla

On the Job-specific settings tab, set the Container Image URI to your Amazon ECR image and configure the scenario parameters (ego speeds, NPC speeds, NPC distances) and camera viewpoints.