Skip to content

Multi-Agent Connected Autonomous Driving (MACAD) Gym environments for Deep RL. Code for the paper presented in the Machine Learning for Autonomous Driving Workshop at NeurIPS 2019:

License

Notifications You must be signed in to change notification settings

praveen-palanisamy/macad-gym

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

MACAD-Gym learning environment 1 MACAD-Gym is a training platform for Multi-Agent Connected Autonomous Driving (MACAD) built on top of the CARLA Autonomous Driving simulator.

MACAD-Gym provides OpenAI Gym-compatible learning environments for various driving scenarios for training Deep RL algorithms in homogeneous/heterogenous, communicating/non-communicating and other multi-agent settings. New environments and scenarios can be easily added using a simple, JSON-like configuration.

PyPI version fury.io PyPI format Downloads

Quick Start

Install MACAD-Gym using pip install macad-gym. If you have CARLA_SERVER setup, you can get going using the following 3 lines of code. If not, follow the Getting started steps.

Training RL Agents

import gym
import macad_gym
env = gym.make("HomoNcomIndePOIntrxMASS3CTWN3-v0")

# Your agent code here

Any RL library that supports the OpenAI-Gym API can be used to train agents in MACAD-Gym. The MACAD-Agents repository provides sample agents as a starter.

Visualizing the Environment

To test-drive the environments, you can run the environment script directly. For example, to test-drive the HomoNcomIndePOIntrxMASS3CTWN3-v0 environment, run:

python -m macad_gym.envs.homo.ncom.inde.po.intrx.ma.stop_sign_3c_town03

Usage guide

Getting Started

Assumes an Ubuntu (18.04/20.04/22.04 or later) system. If you are on Windows 10/11, use the CARLA Windows package and set the CARLA_SERVER environment variable to the CARLA installation directory.

  1. Install the system requirements:

    • Miniconda/Anaconda 3.x
      • wget -P ~ https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh; bash ~/Miniconda3-latest-Linux-x86_64.sh
    • cmake (sudo apt install cmake)
    • zlib (sudo apt install zlib1g-dev)
    • [optional] ffmpeg (sudo apt install ffmpeg)
  2. Setup CARLA (0.9.x)

    3.1 mkdir ~/software && cd ~/software

    3.2 Example: Download the 0.9.13 release version from: Here Extract it into ~/software/CARLA_0.9.13

    3.3 echo "export CARLA_SERVER=${HOME}/software/CARLA_0.9.13/CarlaUE4.sh" >> ~/.bashrc

  3. Install MACAD-Gym:

    • Option1 for users : pip install macad-gym
    • Option2 for developers:
      • Fork/Clone the repository to your workspace: git clone https://github.com/praveen-palanisamy/macad-gym.git && cd macad-gym
      • Create a new conda env named "macad-gym" and install the required packages: conda env create -f conda_env.yml
      • Activate the macad-gym conda python env: source activate macad-gym
      • Install the macad-gym package: pip install -e .
      • Install CARLA PythonAPI: pip install carla==0.9.13

      NOTE: Change the carla client PyPI package version number to match with your CARLA server version

Learning Platform and Agent Interface

The MACAD-Gym platform provides learning environments for training agents in both, single-agent and multi-agent settings for various autonomous driving tasks and scenarios that enables training agents in homogeneous/heterogeneous The learning environments follows naming convention for the ID to be consistent and to support versioned benchmarking of agent algorithms. The naming convention is illustrated below with HeteCommCoopPOUrbanMgoalMAUSID as an example: MACAD-Gym Naming Conventions

The number of training environments in MACAD-Gym is expected to grow over time (PRs are very welcome!).

Environments

The environment interface is simple and follows the widely adopted OpenAI-Gym interface. You can create an instance of a learning environment using the following 3 lines of code:

import gym
import macad_gym
env = gym.make("HomoNcomIndePOIntrxMASS3CTWN3-v0")

Like any OpenAI Gym environment, you can obtain the observation space and action spaces as shown below:

>>> print(env.observation_space)
Dict(car1:Box(168, 168, 3), car2:Box(168, 168, 3), car3:Box(168, 168, 3))
>>> print(env.action_space)
Dict(car1:Discrete(9), car2:Discrete(9), car3:Discrete(9))

To get a list of available environments, you can use the list_available_envs() function as shown in the code snippet below:

import gym
import macad_gym
macad_gym.list_available_envs()

This will print the available environments. Sample output is provided below for reference:

Environment-ID: Short description
{'HeteNcomIndePOIntrxMATLS1B2C1PTWN3-v0': 'Heterogeneous, Non-communicating, '
                                          'Independent,Partially-Observable '
                                          'Intersection Multi-Agent scenario '
                                          'with Traffic-Light Signal, 1-Bike, '
                                          '2-Car,1-Pedestrian in Town3, '
                                          'version 0',
 'HomoNcomIndePOIntrxMASS3CTWN3-v0': 'Homogenous, Non-communicating, '
                                     'Independed, Partially-Observable '
                                     'Intersection Multi-Agent scenario with '
                                     'Stop-Sign, 3 Cars in Town3, version 0'}

Agent interface

The Agent-Environment interface is compatible with the OpenAI-Gym interface thus, allowing for easy experimentation with existing RL agent algorithm implementations and libraries. You can use any existing Deep RL library that supports the Open AI Gym API to train your agents.

The basic agent-environment interaction loop is as follows:

import gym
import macad_gym


env = gym.make("HomoNcomIndePOIntrxMASS3CTWN3-v0")
configs = env.configs
env_config = configs["env"]
actor_configs = configs["actors"]


class SimpleAgent(object):
    def __init__(self, actor_configs):
        """A simple, deterministic agent for an example
        Args:
            actor_configs: Actor config dict
        """
        self.actor_configs = actor_configs
        self.action_dict = {}


    def get_action(self, obs):
        """ Returns `action_dict` containing actions for each agent in the env
        """
        for actor_id in self.actor_configs.keys():
            # ... Process obs of each agent and generate action ...
            if env_config["discrete_actions"]:
                self.action_dict[actor_id] = 3  # Drive forward
            else:
                self.action_dict[actor_id] = [1, 0]  # Full-throttle
        return self.action_dict


agent = SimpleAgent(actor_configs)  # Plug-in your agent or use MACAD-Agents
for ep in range(2):
    obs = env.reset()
    done = {"__all__": False}
    step = 0
    while not done["__all__"]:
        obs, reward, done, info = env.step(agent.get_action(obs))
        print(f"Step#:{step}  Rew:{reward}  Done:{done}")
        step += 1
env.close()

Citing:

If you find this work useful in your research, please cite:

@misc{palanisamy2019multiagent,
    title={Multi-Agent Connected Autonomous Driving using Deep Reinforcement Learning},
    author={Praveen Palanisamy},
    year={2019},
    eprint={1911.04175},
    archivePrefix={arXiv},
    primaryClass={cs.LG}
}
Citation in other Formats: (Click to View)

MLA
Palanisamy, Praveen. "Multi-Agent Connected Autonomous Driving using Deep Reinforcement Learning." arXiv preprint arXiv:1911.04175 (2019).
APA
Palanisamy, P. (2019). Multi-Agent Connected Autonomous Driving using Deep Reinforcement Learning. arXiv preprint arXiv:1911.04175.
Chicago
Palanisamy, Praveen. "Multi-Agent Connected Autonomous Driving using Deep Reinforcement Learning." arXiv preprint arXiv:1911.04175 (2019).
Harvard
Palanisamy, P., 2019. Multi-Agent Connected Autonomous Driving using Deep Reinforcement Learning. arXiv preprint arXiv:1911.04175.
Vancouver
Palanisamy P. Multi-Agent Connected Autonomous Driving using Deep Reinforcement Learning. arXiv preprint arXiv:1911.04175. 2019 Nov 11.

NOTEs:

  • MACAD-Gym supports multi-GPU setups and it will choose the GPU that is less loaded to launch the simulation needed for the RL training environment

  • MACAD-Gym is for CARLA 0.9.x & above . If you are looking for an OpenAI Gym-compatible agent learning environment for CARLA 0.8.x (stable release), use this carla_gym environment.