DI-drive

Introduction

DI-drive doc

DI-drive is an open-source Decision Intelligence Platform for Autonomous Driving simulation. DI-drive applies different simulators/datasets/cases in Decision Intelligence Training & Testing for Autonomous Driving Policy. It aims to

• run Imitation Learning, Reinforcement Learning, GAIL etc. in a single platform and simple unified entry
• apply Decision Intelligence in any part of the driving simulation
• suit most of the driving simulators input & output
• run designed driving cases and scenarios

and most importantly, to put these all together!

DI-drive uses DI-engine, a Reinforcement Learning platform to build most of the running modules and demos. DI-drive currently supports Carla, an open-source Autonomous Driving simulator to operate driving simulation, and MetaDrive, a diverse driving scenarios for Generalizable Reinforcement Learning. DI-drive is an application platform under OpenDILab

Visualization of Carla driving in DI-drive

Outline

• Introduction
• Outline
• Installation
• Quick Start
• Model Zoo
• Casezoo
• File Structure
• Contributing
• License
• Citation

Installation

DI-drive runs with Python >= 3.5 and DI-engine >= 0.3.1 (Pytorch is needed in DI-engine). You can install DI-drive from the source code:

git clone https://github.com/opendilab/DI-drive.git
cd DI-drive
pip install -e .

DI-engine and Pytorch will be installed automatically.

In addition, at least one simulator in Carla and MetaDrive need to be installed to run in DI-drive. MetaDrive can be easily installed via pip. If Carla server is used for simulation, users need to install 'Carla Python API' in addition. You can use either one of them or both. Make sure to modify the activated simulators in core.__init__.py to avoid import error.

Please refer to the installation guide for details about the installation of DI-drive.

Quick Start

Carla

Users can check the installation of Carla and watch the visualization by running an 'auto' policy in provided town map. You need to start a Carla server first and modify the Carla host and port in auto_run.py into yours. Then run:

cd demo/auto_run
python auto_run.py

MetaDrive

After installation of MetaDrive, you can start an RL training in MetaDrive Macro Environment by running the following code:

cd demo/metadrive
python macro_env_dqn_train.py.

We provide detail guidance for IL and RL experiments in all simulators and quick run of existing policy for beginners in our documentation. Please refer to it if you have further questions.

Model Zoo

Imitation Learning

• Conditional Imitation Learning
• Learning by Cheating
• from Continuous Intention to Continuous Trajectory

Reinforcement Learning

• BeV Speed RL
• Implicit Affordance
• Latent DRL
• MetaDrive Macro RL

Other Method

• DREX

DI-drive Casezoo

DI-drive Casezoo is a scenario set for training and testing the Autonomous Driving policy in simulator. Casezoo combines data collected from actual vehicles and Shanghai Lingang road license test Scenarios. Casezoo supports both evaluating and training, which makes the simulation closer to real driving.

Please see casezoo instruction for details about Casezoo.

File Structure

DI-drive
|-- .gitignore
|-- .style.yapf
|-- CHANGELOG
|-- LICENSE
|-- README.md
|-- format.sh
|-- setup.py
|-- core
|   |-- data
|   |   |-- base_collector.py
|   |   |-- benchmark_dataset_saver.py
|   |   |-- bev_vae_dataset.py
|   |   |-- carla_benchmark_collector.py
|   |   |-- cict_dataset.py
|   |   |-- cilrs_dataset.py
|   |   |-- lbc_dataset.py
|   |   |-- benchmark
|   |   |-- casezoo
|   |   |-- srunner
|   |-- envs
|   |   |-- base_drive_env.py
|   |   |-- drive_env_wrapper.py
|   |   |-- md_macro_env.py
|   |   |-- md_traj_env.py
|   |   |-- scenario_carla_env.py
|   |   |-- simple_carla_env.py
|   |-- eval
|   |   |-- base_evaluator.py
|   |   |-- carla_benchmark_evaluator.py
|   |   |-- serial_evaluator.py
|   |   |-- single_carla_evaluator.py
|   |-- models
|   |   |-- bev_speed_model.py
|   |   |-- cilrs_model.py
|   |   |-- common_model.py
|   |   |-- lbc_model.py
|   |   |-- model_wrappers.py
|   |   |-- mpc_controller.py
|   |   |-- pid_controller.py
|   |   |-- vae_model.py
|   |   |-- vehicle_controller.py
|   |-- policy
|   |   |-- traj_policy
|   |   |-- auto_policy.py
|   |   |-- base_carla_policy.py
|   |   |-- cilrs_policy.py
|   |   |-- lbc_policy.py
|   |-- simulators
|   |   |-- base_simulator.py
|   |   |-- carla_data_provider.py
|   |   |-- carla_scenario_simulator.py
|   |   |-- carla_simulator.py
|   |   |-- fake_simulator.py
|   |   |-- srunner
|   |-- utils
|       |-- data_utils
|       |-- env_utils
|       |-- learner_utils
|       |-- model_utils
|       |-- others
|       |-- planner
|       |-- simulator_utils
|-- demo
|   |-- auto_run
|   |-- cict
|   |-- cilrs
|   |-- implicit
|   |-- latent_rl
|   |-- lbc
|   |-- metadrive
|   |-- simple_rl
|-- docs
|   |-- casezoo_instruction.md
|   |-- figs
|   |-- source

Join and Contribute

We appreciate all contributions to improve DI-drive, both algorithms and system designs. Welcome to OpenDILab community! Scan the QR code and add us on Wechat:

Or you can contact us with slack or email (opendilab@pjlab.org.cn).

License

DI-engine released under the Apache 2.0 license.

Citation

@misc{didrive,
    title={{DI-drive: OpenDILab} Decision Intelligence platform for Autonomous Driving simulation},
    author={DI-drive Contributors},
    publisher = {GitHub},
    howpublished = {\url{https://github.com/opendilab/DI-drive}},
    year={2021},
}