Run an autonomous driving simulation sweep with CARLA

This tutorial walks you through running a CARLA autonomous driving simulation parameter sweep with configurable multi-sensor capture. You build a Docker image, push it to Amazon Elastic Container Registry (Amazon ECR), and submit the autonomous driving CARLA job bundle on GitHub to a GPU fleet on your Deadline Cloud farm.

The job runs a lane-change cut-in scenario where an NPC vehicle starts behind the ego vehicle, accelerates to position itself 20 meters ahead of the ego during a 105-second get-ahead phase, then cuts into the ego's lane. It sweeps across configurable ego speeds, NPC speeds, and NPC starting distances, creating a task for each parameter combination (default 2×2×2 = 8 tasks). Each task captures multi-sensor data from the camera viewpoints you select and produces the following output:

The following video demonstrates the CARLA simulation sweep workflow on Deadline Cloud.

Estimated time: 1–2 hours, including the Docker image build. Tasks typically complete in about 13 minutes each.

Running this tutorial incurs charges for the GPU worker instances that process the tasks and for Amazon ECR image storage.

Overview

To complete this tutorial, follow these steps:

Prerequisites

Before you begin, complete the following setup:

Set up your farm

You need a Deadline Cloud farm with a queue that has a conda queue environment (channels: conda-forge, packages: ffmpeg) and a GPU fleet (minimum: 1 NVIDIA GPU with 16 vCPU and 64 GiB memory).

IAM permissions – Your queue role needs Amazon ECR pull permissions, because the task script runs docker pull under queue role credentials. Attach a policy like the following:

{
  "Version": "2012-10-17",
  "Statement": [
    {
      "Effect": "Allow",
      "Action": [
        "ecr:GetDownloadUrlForLayer",
        "ecr:BatchGetImage",
        "ecr:BatchCheckLayerAvailability"
      ],
      "Resource": "arn:aws:ecr:REGION:ACCOUNT_ID:repository/*"
    },
    {
      "Effect": "Allow",
      "Action": "ecr:GetAuthorizationToken",
      "Resource": "*"
    }
  ]
}

Fleet host configuration – Your fleet workers need Docker and the NVIDIA Container Toolkit. Attach the Docker and NVIDIA Container Toolkit host configuration script on GitHub to your fleet. For more information, see Run Docker containers with NVIDIA GPUs on Deadline Cloud workers.

Build and push the Docker image

The job runs inside a Docker container based on the carlasim/carla:0.9.16 base image on Docker Hub. The base CARLA image includes the simulator but lacks the Python environment, scenario runner, and sensor capture scripts needed for this job. The custom image layers on Python 3.10, scenario_runner, and the entrypoint and capture scripts so each Deadline Cloud task can boot CARLA, execute the driving scenario, and record sensor data in a single container.

Submit the simulation job

Alternatively, submit with the CLI:

deadline bundle submit . \
  --farm-id FARM_ID \
  --queue-id QUEUE_ID \
  --name "CARLA Lane Change Demo" \
  -p ImageURI=ACCOUNT_ID.dkr.ecr.REGION.amazonaws.com/carla-deadline:0.9.16 \
  -p AwsRegion=REGION

Monitor the job

Monitor progress in the Deadline Cloud console. Each task shows its parameter values (EgoSpeed, NpcSpeed, NpcDistance) in the task table. Tasks typically complete in about 13 minutes each.

The log output shows:

Review the output

Each task produces output in a subdirectory named for its parameters:

outputs/
└── ego20_npc30_dist10/
    ├── rgb/
    │   ├── front/frame_000001.png ... frame_000062.png
    │   └── rear/frame_000001.png ... frame_000062.png
    ├── semantic/
    │   ├── front/...
    │   └── rear/...
    ├── lidar/frame_000001.ply ...
    ├── bbox_2d/{front,rear}/frame_*.txt
    ├── bbox_3d/frame_*.txt
    └── video/
        ├── front_scenario.mp4
        ├── rear_scenario.mp4
        └── grid_scenario.mp4

Known limitations

Clean up

To avoid ongoing charges, clean up the resources that you created for this tutorial:

Related resources

The following resources provide additional information: