DeepRacer notebook using Amazon SageMaker RL and AWS RoboMaker services
This folder contains examples of how to use RL to train an autonomous deepracer. This is a jailbreaker for the AWS DeepRacer. This gives a glimse of architecture used to get the DeepRacer working.
Contents
-
deepracer_rl.ipynb
: notebook for training autonomous race car. -
Dockerfile
: Custom docker instead of using SageMaker default docker -
src/
training_worker.py
: Main entrypoint for starting distributed training jobmarkov/
: Helper files for S3 upload/downloadpresets/default.py
: Preset (configuration) for DeepRacerrewards/default.py
: Custom reward functionenvironments/deepracer_racetrack_env.py
: Gym environment file for DeepRaceractions/model_metadata_10_state.json
: JSON file to customize your action space & the speedlib/
: redis configuration file and ppo_head.py customized tensorflow file copied to sagemaker container.
-
common/
: helper function to build docker files.
How to use the notebook
- Login to your AWS account - SageMaker service (SageMaker Link)
- On the left tab select
Notebook instances
- Select
Create notebook instance
- Fill up the notebook instance name. In the Additional configuration select atleast 25GB. This is because docker gets installed and takes up space.
- Create a new IAM role. Give root permission
- Select the git repository and clone this repository.
- Then click create notebook instance button at the button
- This takes like 2 min to create your notebook instance. Then click on the newly created instance and click on the juypter notebook.
- You will see all the github files and now run
deepracer_rl.ipynb
- Run clean robomaker & sagemaker commands in the script only when you are done with training.
DeepRacer Paper
A techncial paper for AWS DeepRacer is available at https://arxiv.org/abs/1911.01562. Below is a BibTeX entry for citations:
@misc{deepracer2019,
title={DeepRacer: Educational Autonomous Racing Platform for Experimentation with Sim2Real Reinforcement Learning},
author={Bharathan Balaji and Sunil Mallya and Sahika Genc and Saurabh Gupta and Leo Dirac and Vineet Khare and Gourav Roy and Tao Sun and Yunzhe Tao and Brian Townsend and Eddie Calleja and Sunil Muralidhara and Dhanasekar Karuppasamy},
year={2019},
eprint={1911.01562},
archivePrefix={arXiv},
primaryClass={cs.LG}
}