Highway: A light and open-source simulator for autonomous driving research.

Note: this repo is given 'as it' since I'm currently have no time to work on it. There are lot of things to do before having a complete simulator.

Why another vehicle simulator ?

Fig 1 - A parallel automatic maneuver. Click on the image to run the video.

This simulator has not the pretention to revolutionize the world of the vehicle simulation but is just a personal challenge for offering an intuitive API for autonomous driving research.

When I tried Carla simulator, I did not like the following points:

• Hard to install (Unreal Engine, NVidia drivers, patches, need Docker ...) and takes giga bytes of binaries on the hard disk. The application is low to run.
• Very instable: lot of crashes, lot of issues on GitHub that are never fixed and a robot is auto-closing them after X days without answers.
• The physical models of cars are not very realist: looks like playing the GTA III game.
• The most important is the API is not developper friendly and poorly documented. For example, When reading the first lines of the API: carla.Actor: defines actors as anything that plays a role in the simulation or can be moved around. That includes: pedestrians, vehicles, sensors and traffic signs ... but when reading other Actor methods we reach close_door() ... We can interrogate if this API has been really thought.
• The documentation is sometimes hard to understand. I was personnaly stuck searching on the API what methods giving me dimensions of my car (wheel base ...), or how to understand how they have implemented Akermann ...

In my opinion, a simulator has to help the developper by offering the possibility to interact with all objects in the word. For example the simulator has to reply to question such as "give me the list of pedestrians around the car" ... a reference list that can be compared at run time with the list of pedestrians detected by your IA/sensors/ECU of your ego vehicle. This idea can be extended to any mobile/static object in the city: pedestrian, cars, roads, traffic signs, zebras, parking slot, spawning points ... all of them shall interact with the ego vehicle and share their states (ie traffic light: the color of the light, the content of any road signs ie "Nationale 20" ...).

Ideally, the API shall distinguish helper methods to twick the simulation from real physical behavior by explicit methods for example apply a constant velocity along its longitudinal side.

Here a list of things a simulator can help and that I did not find in Carla:

• Triggering collisions events;
• Offering to the developper different physical models (kinematics, dynamics ...) that can be changed at run time or offering to the player the way to develop it's own model. Same idea for the cruse controller.
• Accessing to blueprint/dimension of the car (wheelbase ...), parking, roads, objects around;
• Helping the developper cheating: for example, applying a constant longitudinal velocity to the car if we do not want to simulate the brake and the accelerator pedals;
• Monitoring vehicle states and save logs that be replayed or plot graphics (for example in Matlab or Julia);
• Displaying shapes to inform the developper: for example the arc for turning circle, show spawning points identifier, show arrows for the vehicle speed, acceleration, the car path ...

Compilation / Installation

Compilation

This code depends on:

• C++-17 compiler (g++, clang++)
• SFML for 2D graphism: sudo apt-get install libsfml-dev.
• SWI Prolog sudo apt-get install swi-prolog. Prolog is a logic programming language associated with artificial intelligence and computational linguistics. Currently it is not used.
• Backward-cpp for stack tracing on segfaults. This lib is automatically downloaded (but not installed on your operating system) when compiling in debug mode.
• Google tests for unit tests: https://github.com/google/googletest.

To compile and install the standalone application:

• Download third part librairies. This has to be call once:

git clone git@github.com:Lecrapouille/Highway.git --recursive --depth=1
make download-external-libs
make compile-external-libs

Once you can compile the project.

make -j8

If you prefer instead to choose a different compiler, i.e:

make CXX=clang++-13 -j8

Scenario files are automatically compiled. You have to do nothing. Here what is made (hidden):

for i in `ls Scenarios` do
  (cd $i
   make clean
   make -j8)
end
cp Scenarios/*/build/*.so data/Scenarios/

The compiled simulation files are in Scenarios/xxx/build/xxx.so but then are copied into the data/Scenarios folder to be loaded by the application. There are pure shared libraries that can be opened by with dlopen() function.

To run the standalone application.

./build/Highway

To run the standalone application and running a given scenario (shared library). This will skip some IHM menu and jump directly into the simulation.

Highway scenario.so

To install the application and scenarios on your operating system:

sudo make install

To run the standalone application installed on your system:

Highway

To clean the project:

make clean

This will remove the buid/ folder. To deep clean (third parts, docs ...) type make veryclean.

Documentation

• The index of detail design documents: doc/README.md.

• The code source refers in comments to mathematic formulas from the documents cited inside bibliography sections on each folders. Formulas are pictures. For Emacs users, when diving inside the code source, you can type M-x turn-on-iimage-mode to show pictures inside C++ comments. This will help you to understand the code bind to mathematic formulas of these documents.

• To generate Doxygen documentation.

make doc

The generated documentation can be find in doc/html.

Developpers

Non regression tests

To make unit tests and generate code coverage.

make check -j8

Or

cd tests
make coverage -j8

If you do not want generating reports.

cd tests
make -j8
./build/Highway-Tests

Compress your work

To create a tar.gz (for a backup) of the project with management of name conflict concerning the tarball name. Compiled files, generated doc, git files and backup files are not stored in the tarball.

make tarball