DeepRacer is back - First impressions of "DeepRacer on AWS"!
The wait is over. The much awaited "DeepRacer on AWS" Solution is out - it is now possible to deploy the entire DeepRacer stack into your own AWS account! I have been taking it for a spin.
Around re:Invent 2024, AWS announced that the DeepRacer service would be decommissioned on December 15, 2025, and that a new solution would be available by then to allow individuals and organizations to deploy DeepRacer into their own AWS accounts. December 15 came and went, and no solution was in sight. But today is the day - the new DeepRacer on AWS solution is available in the AWS Solutions Library!
Solution vs. Service
AWS Services are the foundational components of AWS (like S3, EC2, CloudWatch, etc.) and are managed through the AWS Console or CLI. They typically don’t provide end‑user functionality. DeepRacer was different: It was an end‑user product hosted in the AWS Console, built on top of services like SageMaker, RoboMaker, Lambda, S3, CloudWatch, Kinesis, and others.
A major pain point with the DeepRacer service was that every racer needed AWS Console access. For individuals this wasn’t a big issue, but for larger corporate events it was a nightmare. Integration with corporate identity and access management was possible, but cumbersome for both admins and racers.
A Solution, in contrast, is an application provided by AWS that you deploy into your own AWS account. You can run it as‑is or modify it because the source code is available. Best of all, it typically provides a normal website with user access managed by Cognito.
First Impressions
Installation
I decided to install the out-of-the box solution into my private AWS account.Like other AWS Solutions, it comes with a solid guide offering different deployment options ; the simplest option is deploying a CloudFormation template. (Click here for instant deployment !)
As documented, the CloudFormation installation takes about 30 minutes. At the end, you receive an email with the URL and password. After logging in, you’re greeted by a basic welcome screen.
Next Steps
Invite users:
The logical next step is to go to Manage Instance and invite another user. By default, the registration mode is Invite only, and I didn’t find a way to change this. To invite a user by email, go to Users → Actions → Invite user. The user receives an email with a link and initial password and can then define their Racer Name/Alias. Manage Instance also provides basic usage quota options (training/evaluation time and number of models).
The logical next step is to go to Manage Instance and invite another user. By default, the registration mode is Invite only, and I didn’t find a way to change this. To invite a user by email, go to Users → Actions → Invite user. The user receives an email with a link and initial password and can then define their Racer Name/Alias. Manage Instance also provides basic usage quota options (training/evaluation time and number of models).
Create a race:
Creating a race is nearly identical to creating a community race in the Console. Since those screens were already well‑designed, “less is more” applies here.
Creating a race is nearly identical to creating a community race in the Console. Since those screens were already well‑designed, “less is more” applies here.
Train a model:
Logging in as a racer, go to Learning and Models → Your Models → Create Model. You’ll find a wizard very similar to the Console’s model creation flow. Step through the process as before, and you’ll reach the Model page showing Training, Evaluation, and Training Configuration information — just like you’re used to. You can even submit to the race created earlier.
Logging in as a racer, go to Learning and Models → Your Models → Create Model. You’ll find a wizard very similar to the Console’s model creation flow. Step through the process as before, and you’ll reach the Model page showing Training, Evaluation, and Training Configuration information — just like you’re used to. You can even submit to the race created earlier.
Features
Most of the core Console features have been retained. The main screens (create race, create model, model view) are almost identical, with the main visible difference being the use of the Cloudscape Design System , giving everything a more modern look.
Beyond the main screens, the functionality is fairly basic — similar to Community Races, but without the leaderboard functionality of the Virtual League. Country/regional affiliation for racers is also gone.
Verdict
The Good
- It works. Deployment is easy. No compilation required.
- Main screens are familiar.
- It’s cheaper: training time is now billed hourly for SageMaker ml.c5.4xlarge, around $0.83 in us‑east‑1.
- It’s open: with access to the source, it should be possible to fix known DeepRacer Service issues, such as the lack of GPU support that makes training painfully slow.
The Bad
- It still runs on SageMaker. Any job (training or evaluation) takes 5+ minutes to start.
- Evaluations requiring one minute of driving can still run for up to 10 minutes.
- CPU-only training is still slow.
- Live Racing is missing.
The Ugly
- You can’t view race submission videos, which is strange. Racing is dull if you can’t see how others are performing.
The Future
Overall, this is exactly what the community has been waiting for. While basic in some areas, it provides everything needed to train models and let groups race virtually — or combine it with the DeepRacer Event Manager for physical racing.
To make it great, a few things are needed:
- Accelerated training using GPU and multiple workers (as known from DeepRacer‑for‑Cloud), enabling workshop participants to train more advanced models faster.
- Integration with DREM — a shared username/password plus a model‑transfer feature would dramatically improve workshop user experience.
DeepRacer was dead. Long live DeepRacer! So what are you waiting for?
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article