| URL | |
|---|---|
| Github | https://github.com/tier4/AWML |
| arXiv | https://arxiv.org/abs/2506.00645 |
This repository is a machine learning library to deploy for Autoware to aim for "robotics MLOps".
AWML supports training with T4dataset format in addition to open datasets and deployment for Autoware.
In addition to ML model deployment, AWML supports active learning framework include auto labeling, semi-auto labeling, and data mining.
AWML can deploy following task for now.
- 2D detection for dynamic recognition
- 3D detection for dynamic recognition
- 2D fine detection for traffic light recognition
- 2D classification for traffic light recognition
AWML supports following environment.
- All tools are tested by Docker environment on Ubuntu 22.04LTS
- NVIDIA dependency: CUDA 12.1 + cuDNN 8
- Need > 530.xx.xx NVIDIA device driver
If you use this project in your research, please use the arXiv paper and consider cite.
@misc{tanaka2025awmlopensourcemlbasedrobotics,
title={AWML: An Open-Source ML-based Robotics Perception Framework to Deploy for ROS-based Autonomous Driving Software},
author={Satoshi Tanaka and Samrat Thapa and Kok Seang Tan and Amadeusz Szymko and Lobos Kenzo and Koji Minoda and Shintaro Tomie and Kotaro Uetake and Guolong Zhang and Isamu Yamashita and Takamasa Horibe},
year={2025},
eprint={2506.00645},
archivePrefix={arXiv},
primaryClass={cs.RO},
url={https://arxiv.org/abs/2506.00645},
}
AWML is based on Autoware Core & Universe strategy.
We hope that AWML promotes the community between Autoware and ML researchers and engineers.
| URL | |
|---|---|
| Autoware | https://github.com/autowarefoundation/autoware |
| AWMLPrediction | TBD |
If you want to know about the design of AWML, you should read following pages.
- Docs for architecture of dataset pipeline
- Docs for architecture of ML model
- Docs for architecture of S3 storage
- Docs for architecture of evaluation
- Docs for AWML design
- Docs for deploy flow design
If you want to develop AWML, you should read following pages.
If you want to search the OSS tools around AWML, you should read following pages.
If you want to know about AWML, you should read following pages.
- Setting environment for AWML
- Training, evaluation, and deployment
- Analyze for the dataset and the model
- Auto labeling
- Data mining
- ROS2
- Supported models
- ⭐ is recommended to use
| Task | Model | Use for Autoware |
|---|---|---|
| 3D detection | CenterPoint | ⭐ |
| 3D detection | TransFusion | ✅ |
| 3D detection | BEVFusion | ✅ |
| 3D segmentation | FRNet | (Reviewing now) |
| 2D detection | YOLOX | |
| 2D detection | YOLOX_opt | ⭐ |
| 2D detection | GLIP | |
| 2D detection | SwinTransformer | |
| 2D classification | MobileNetv2 | ✅ |
| Vision language | BLIP-2 | |
- Additional plug-ins