aws-deepracer-community/deepracer-for-cloud

What it solves

It provides a streamlined way to set up and run AWS DeepRacer training environments on your own hardware or cloud virtual machines (such as AWS EC2 or Azure N-Series), removing the dependency on the official AWS DeepRacer console service.

How it works

The project uses a set of scripts and Docker containers to orchestrate a robotics simulation environment. It leverages the DeepRacer Simapp image, which combines three core components: Robomaker (for ROS and Gazebo simulation), Sagemaker (for the model training job), and RL Coach (to bootstrap the training process). It supports both Docker Compose and Docker Swarm for managing these containers, allowing users to scale training across multiple nodes.

Who it’s for

It is designed for DeepRacer users who want more control over their training environment, those who wish to avoid the constraints of the official AWS console, or those looking to optimize training speed and cost using their own GPU or CPU-powered cloud instances.

Highlights

• Flexible Hardware Support: Works on Ubuntu, macOS (via Colima), and supports NVIDIA GPUs, Intel CPUs, and ARM64/Graviton processors.
• Scalable Training: Supports running multiple simulation instances per training job (N:1) and parallel training sessions to speed up progress.
• Multi-Node Clusters: Ability to connect multiple computers via Docker Swarm to combine computing power.
• Comprehensive Evaluation: Allows for independent model evaluation and the ability to save evaluation runs as MP4 files to S3.
• AWS Integration: Can detect AWS DeepRacer Console models and upload locally trained models back to them.