ScenarioRunner for CARLA

This repository contains traffic scenario definition and an execution engine for CARLA. It also allows the execution of a simulation of the CARLA Challenge. You can use this system to prepare your agent for the CARLA Challenge.

Scenarios can be defined through a Python interface, and with the newest version the scenario_runner also the upcoming OpenScenario standard is supported.

Getting the ScenarioRunner

Use git clone or download the project from this page. Note that the master branch contains the latest fixes and features, and may be required to use the latest features from CARLA.

It is important to also consider the release version that has to match the CARLA version.

• Version 0.9.6: Compatible with CARLA 0.9.6
• Version 0.9.5 and Version 0.9.5.1: Compatible with CARLA 0.9.5
• Version 0.9.2: Compatible with CARLA 0.9.2

To use a particular version you can either download the corresponding tarball or simply checkout the version tag associated to the release (e.g. git checkout v0.9.5)

Currently no build is required, as all code is in Python.

Using the ScenarioRunner

Please take a look at our Getting started documentation.

Challenge Evaluation

You can evaluate your own agents using a reproduction of the CARLA Challenge by following this tutorial

Contributing

Please take a look at our Contribution guidelines.

F.A.Q.

If you run into problems, check our FAQ.

License

ScenarioRunner specific code is distributed under MIT License.

Customize ScenarioRunner for MPSC

Requirement (Installation)

1. Install UnrealEngine4.22 and Latest CARLA following [How to build CARLA on Linux].
   Note:

• clone master branch instead of stable branch to aviod "No rule to build target 'launch'" error. (Two branches have totally different Makefile)
• They just updated the Releases to 0.9.6, but I am using 0.9.5.

2. Clone this repo and follow Getting started documentation.
3. Python3.5

NN_controller

Run the code in scenario_runner setting

1. After [building] and Getting started and some configurations, my ~/.bashrc looks like the followings:

# >>> for CARLA >>>
# start UE as CARLA's server
export UE4_ROOT=~/UnrealEngine_4.22
 
export CARLA_ROOT=/home/ruihan/UnrealEngine_4.22/carla/
export PYTHONPATH=${CARLA_ROOT}/PythonAPI/carla:$PYTHONPATH

export CARLA_SERVER_DIR=/home/ruihan/UnrealEngine_4.22/carla/Unreal/CarlaUE4/Saved/StagedBuilds/LinuxNoEditor
# export CARLA_SERVER=/home/ruihan/UnrealEngine_4.22/carla/Unreal/CarlaUE4/Saved/StagedBuilds/LinuxNoEditor/CarlaUE4.sh

# for conditional imitation learning
export PYTHONPATH=/home/ruihan/coiltraine:$PYTHONPATH

# for CARLA AD Challenge, scenario_runner
export ROOT_SCENARIO_RUNNER=/home/ruihan/scenario_runner
export PYTHONPATH=$PYTHONPATH:${CARLA_ROOT}/PythonAPI/carla/dist/carla-0.9.5-py3.5-linux-x86_64.egg:${CARLA_ROOT}/PythonAPI/carla/agents:${CARLA_ROOT}/PythonAPI/carla
export PYTHONPATH=$PYTHONPATH:${CARLA_ROOT}/PythonAPI/carla/:${ROOT_SCENARIO_RUNNER}:${CARLA_ROOT}/PythonAPI/carla/dist/carla-0.9.5-py2.7-linux-x86_64.egg:${CARLA_ROOT}/PythonAPI/carla/agents

(coiltraine environment is not necessary if all other required packages are installed)

2. Run the commands

• To collect data from manual_control.py, press P on keyboard to enable autopilot
  modify ~L265 (about line 265) def write_in_csv(location, waypoint_tf, lane_width, control, ds='localization_relative_coords_ds.csv') to change dataset filename.
  Add record_dataset and write_in_csv methods to original manual_control.py

ruihan@depend-XPS-8930:cd $CARLA_SERVER_DIR
ruihan@depend-XPS-8930:~/UnrealEngine_4.22/carla/Unreal/CarlaUE4/Saved/StagedBuilds/LinuxNoEditor$ ./CarlaUE4.sh
(coiltraine) ruihan@depend-XPS-8930:~/scenario_runner$ python scenario_runner.py --scenario BackgroundActivity_1 --reloadWorld
(coiltraine) ruihan@depend-XPS-8930:~/scenario_runner$ python manual_control_record.py 

• To train the NN_controller ruihan@depend-XPS-8930:~/scenario_runner$ python NN_controller.py
  modify ~L58 ds_file = 'localization_relative_coords_ds.csv' (input dataset) and ~L158 torch.save(model, 'models/NN_model_relative_epo50.pth') (output model)
• To test the NN_controller, press N on keyboard to enable NN_controller
  Add get_nn_controller method to manual_control_record.py

ruihan@depend-XPS-8930:~/UnrealEngine_4.22/carla/Unreal/CarlaUE4/Saved/StagedBuilds/LinuxNoEditor$ ./CarlaUE4.sh
(coiltraine) ruihan@depend-XPS-8930:~/scenario_runner$ python scenario_runner.py --scenario BackgroundActivity_1 --reloadWorld
(coiltraine) ruihan@depend-XPS-8930:~/scenario_runner$ python manual_control_test_NN.py 

Modify ~L118 def __init__(self, carla_world, hud, nn_model_path='models/NN_model_epo50.pth'): to change the path of the model to be tested.

• To see the performance of CoIL

ruihan@depend-XPS-8930:~/UnrealEngine_4.22/carla/Unreal/CarlaUE4/Saved/StagedBuilds/LinuxNoEditor$ ./CarlaUE4.sh -benchmark -fps=20 -quality-level=Epic
(coiltraine) ruihan@depend-XPS-8930:~/scenario_runner$ bash srunner/challenge/run_evaluator_co.sh

Run the code in CARLA AD Challenge setting

ruihan@depend-XPS-8930:~/UnrealEngine_4.22/carla/Unreal/CarlaUE4/Saved/StagedBuilds/LinuxNoEditor$ ./CarlaUE4.sh -benchmark -fps=20 -quality-level=Epic
(coiltraine) ruihan@depend-XPS-8930:~/scenario_runner$ bash srunner/challenge/run_evaluator_NN.sh

e2c_controller

Run the code

1. Collect data in senario_runner setting

ruihan@depend-XPS-8930:~/UnrealEngine_4.22/carla/Unreal/CarlaUE4/Saved/StagedBuilds/LinuxNoEditor$ ./CarlaUE4.sh
(coiltraine) ruihan@depend-XPS-8930:~/scenario_runner$ python scenario_runner.py --scenario BackgroundActivity_1 --reloadWorld
(coiltraine) ruihan@depend-XPS-8930:~/scenario_runner$ python manual_control_record_e2c.py 

collected data (image .png and control .npy) are saved in /data folder but are not uploaded to remote repo

2. Parse data and train the model (coiltraine) ruihan@depend-XPS-8930:~/scenario_runner$ python e2c_controller.py OR e2c_controller_cat.py (which concatenate measurement to latent vector), which load the /data/processed.pkl file as data source.

3. Note: modify save_dir in manual_control_record_e2c.py => _parse_image_and_save, ds_dir and model_path in e2c_controller_cat.py, corespondingly.