View on GitHub

deepracer-for-cloud

Creates an AWS DeepRacing training environment which can be deployed in the cloud, or locally on Ubuntu Linux, Windows or Mac.

DeepRacer on AWS (DRoA) Integration

DeepRacer on AWS is the community-hosted replacement for the original AWS DeepRacer console. DRfC includes a set of droa-* commands that let you manage models in your DRoA installation directly from the command line.

Prerequisites

Install DRoA

Follow the DeepRacer on AWS installation guide to deploy DRoA into your own AWS account.

Configure DRfC

In system.env set:

DR_DROA_URL=https://<your-droa-domain>   # e.g. https://deepracer.aws.example.com
DR_DROA_USERNAME=<your-droa-email>

DR_DROA_URL is the base URL of your DRoA deployment. At runtime, DRfC fetches <DR_DROA_URL>/env.js to discover the region, Cognito pools, API endpoint, and upload bucket automatically — no additional AWS config required.

Security: never store your DRoA password in system.env. All commands prompt for it interactively (or accept --password on the CLI). Credentials are cached in ~/.droa-cache/ for the duration of the session token.

Python environment

Run bin/prepare.sh to create the .venv virtual environment and install the required Python packages (boto3, pyyaml, requests, deepracer-utils). After source bin/activate.sh the venv is active and all droa-* commands are available.

Commands

droa-list-models

List all models in your DRoA installation, sorted newest-first.

droa-list-models [--json]

Output columns: modelId, name, status, trainingStatus, createdAt.

Status Meaning
IMPORTING Import in progress
READY Available for evaluation
TRAINING Training job running
ERROR Import or training failed
DELETING Deletion in progress

droa-get-model

Show details of a single model.

droa-get-model <modelId> [--verbose] [--summary] [--json]
Flag Description
(none) Identity, car config, training config, metadata
--verbose Adds action space and reward function source
--summary Adds mean training metrics (reward, progress) via DeepRacer Utils
--json Raw JSON output

droa-download-logs

Download training or evaluation logs for a model.

droa-download-logs <modelId> [--asset-type TRAINING_LOGS|EVALUATION_LOGS|PHYSICAL_CAR_MODEL|VIRTUAL_MODEL|VIDEOS]
                              [--evaluation-id <id>]
                              [--output <file>]
                              [--summary]
Flag Description
--asset-type Asset type (default: TRAINING_LOGS)
--evaluation-id Required when --asset-type EVALUATION_LOGS
--output / -o Output file path (default: derived from the presigned URL filename)
--summary Print DeepRacer Utils stability summary after download (TRAINING_LOGS only)

The command polls until the asset is ready (up to 5 minutes for VIRTUAL_MODEL).

droa-delete-model

Delete a model. Only models with status READY or ERROR can be deleted.

droa-delete-model <modelId> [-y/--yes]

Without --yes, you are shown the model name and status and must type the model name to confirm. Deletion is asynchronous — the model transitions to DELETING status.

droa-import-model

Import a locally trained DRFC model into DRoA.

droa-import-model (--model-prefix <prefix> | --model-dir <dir>)
                  [--model-name <name>]
                  [--model-description <text>]
                  [--best | --checkpoint <step>]

Source options

Option Description
--model-prefix Pull directly from local MinIO S3 (DR_LOCAL_S3_BUCKET). Defaults --model-name to the prefix.
--model-dir Use a pre-assembled local directory containing all required model files.

Checkpoint selection (--model-prefix only)

Flag Behaviour
(none) Use the last checkpoint
--best Use the best checkpoint
--checkpoint STEP Use the checkpoint at the given training step

What happens

  1. Model files are pulled from MinIO (path-style S3, using DR_MINIO_URL and DR_LOCAL_S3_PROFILE).
  2. training_params.yaml is copied from the bucket (training_params_1.yaml preferred for multi-worker runs). If missing, it is generated from DR_* environment variables.
  3. WORLD_NAME direction suffixes (_cw, _ccw) are stripped and TRACK_DIRECTION_CLOCKWISE is added — required by DRoA's track validation.
  4. Files are uploaded to the DRoA S3 transit bucket and the import API is called.

Required files (when using --model-dir)

Environment variables reference

Variable Location Description
DR_DROA_URL system.env Base URL of your DRoA deployment
DR_DROA_USERNAME system.env DRoA login email
DR_MINIO_URL system.env MinIO endpoint URL (e.g. http://minio:9000)
DR_LOCAL_S3_PROFILE system.env boto3 AWS profile name for MinIO access
DR_LOCAL_S3_BUCKET run.env Local S3 bucket name
DR_LOCAL_S3_MODEL_PREFIX run.env Default model prefix for --model-prefix

All droa-* commands also accept --url, --username, and --password flags to override the environment variables.