DriveSingularity

A scalable multi-agent learning simulator for autonomous driving

Installation

Step 1: install the requirements

For Ubuntu,

[sudo] apt install git build-essential cmake python3 python3-pip libjsoncpp-dev
pip3 install -r requirements.txt

For Mac OS X,

brew install cmake python3 jsoncpp
pip3 install -r requirements.txt

Step 2: Clone the project, build and install

git clone --recursive https://github.com/KornbergFresnel/DriveSingularity.git
cd DriveSingularity
mkdir build
cd build
cmake ..
make
cd ..

Step 3: install pyds

pip3 install .

Usage

Load scenario

DriveSingularity supports custom scenario registration, so you need to load configuration from local scenario configuration file before generating environment. There is a default scenario configuration file locates at DriveSingularity/configs/simple_env_config.json.

import os.path as osp
from pyds import utils

BASE_DIR = ${WORKSPACE}/
CONFIG_BACKUP = osp.join(BASE_DIR, "configs")

env_config = utils.load_env_config(CONFIG_BACKUP, "simple_env_config.json")

For more details on scenario configuration, please refer to Scenario configuration

Generate environment

Environment generation requires only one command:

from pyds import scenarios

env = engine.Environment(env_config,
                             scenario=scenarios.load(args.scenario)())

Then, you can run it as Gym environments:

import numpy as np

done = False
state_dict = env.reset()

while not done:
    action_dict = ...
    next_state_dict, reward_dict, done_dict, _ = env.step(action_dict)
    done = np.alltrue(list(done_dict.values()))
    # ...

Learning with RLLib

The default training support is RLLib. To enable the interaction between engine and the training toolkit, we offer a multiprocess environment wrapper: DriveSingularity/pyds/multiprocess_wrapper.py, you can get more details from the training example: DriveSingularity/examples/multiprocess_train.py

Render: execute rollout.py

After you have completed the training, you can execute the examples/rollout.py to get the render files, then you can use DSRender to replay the render files.

Scenario configuration

A simple example:

{
    "world": "config/resoureces/osm.json",
    "env_name": "simple_osm",
    "render_dir": "render/osm",
    "n_agents": 2,
    "total_vehicles": 40,
    "length_range": [15, 1],
    "width_range": [8, 1],
    "height_range": [15, 3],
    "velocity_range": [0, 5],
    "target_velocity_range": [10, 10],
    "halfLengthRange": [12, 14],
    "halfWidthRange": [7, 8],
    "velocityRange": [10, 20],
    "targetVelocityRange": [30, 50]
}

Environment macro settings

• world: relative path of road map, must be json format.
• env_name: environment name for registery.register_env.
• render_dir: specify the relative path for the rendering result storage.
• n_agents: specify the maximum number of agents.
• total_vehicles: specify the maximum number of agents + social vehicles.

Configuration for agent vehicles

• length_range: for agent, specify the length of agent vehicle, left value denotes the minimum, right value denotes the range, i.e, the range of length will be: [left vaue, left value + right value].
• width_range: for agent, specify the width of agent vehicle, left value denotes the minimum, right value denotes the range, i.e, the range of width will be: [left vaue, left value + right value].
• height_range: useless (currently).
• velocity_range: for agent, specify the initialized speed, left value denotes the initialized speed, left value denotes the minimum, right value denotes the range, i.e, the range of initialized speed will be: [left vaue, left value + right value].
• target_velocity_range: for agent, specify the target speed if no extra interruption (i.e. agent actions), left value denotes the minimum, right value denotes the range, i.e, the range of target speed will be: [left vaue, left value + right value].

Configuration for social vehicles

• halfLengthRange: for social vehicles, specify the half length, with the given range, the length of social vehicle will be [left value * 2, right value * 2].
• halfWidthRange: for social vehicles, specify the half width, with the given range, the width of social vehicle will be [left value * 2, right value * 2].
• velocityRange: for social vehicles, specify the start speed, with the given range, the initialized speed of social vehicle will be [left value * 2, right value * 2].
• targetVelocityRange: for social vehicles, specify the target speed, with the given range, the target speed of social vehicle will be [left value * 2, right value * 2]