Complex Urban Scenarios for Autonomous Driving

1 Introduction

This repository contains codes for the Complex Urban Scenarios for Autonomous Driving under intersection scenarios. In this repository, we deploy a set of autonomous agents developed based on Reinforcement learning(RL) methods.

2 Get started

2.0 system requirements

Ubuntu 16.04-18.04
CARLA 0.9.10.1 and above

2.1 set-up environment

Anaconda is suggested for python environment management. First of all, create the conda env with
conda env create -f gym_carla.yml -n gym_carla then a conda env named with gym_carla will be created, the conda env can be activated with command
conda activate gym_carla

Before you do anything, please make sure the carla client is running correctly, following carla commands are suggested: ./CarlaUE4.sh -opengl -quality-level=Low -ResX=400 -ResY=300 -carla-rpc-port=2000 Carla port number is alternative, however you should make sure it coordinates with the value in the codes.

Fix the carla root path to get the available carla python API by fixing
./train/gym_carla/config/carla_config.py
Replace the root_path with your own path.

2.1 training

Entrance of the training can be found in
./train/rl_agents/td3/single_task_without_attention

You can train the RL agent with specified task by running rl_agent.py
Some parameters are provided for more training set-ups.
Use python rl_agent.py --help to learn the usage of available parameters.

During the training procedure, checkpoints and best models will be saved in ./outputs/.

We provide a baseline agent with trained weights stored in
train/rl_agents/challenge_agents/test_agents
You can replace the weight file in model_dict folder with your own trained model.

2.2 Testing

2.2.1 Running the test

In order to test trained agent in testing scenarios, please run the ./test/scenario_runner/run_scenario_test.py An agent file is supposed to be assigned, for example, you can run the script with following args:

python run_scenario_test.py
--sync
--reloadWorld
--timeout 99999
--port 2000
--trafficManagerPort 8100
--scenario TestContinuous-right-negative_x-straight
--agent $Your parent folder$/ComplexUrbanScenarios/train/rl_agents/challenge_agents/test_agents/test_agent.py
--speed-range 10 12 2
--distance-range 34 36 2
--tag test

2.2.2 Testing your model

We deploy a set of traffic scenarios for testing the trained agent. If you wish to test your trained agent, please put your tensorflow NN weights in
train/rl_agents/challenge_agents/test_agents/model_dict
Model dict should be put in correct folder of different task.

2.2.3 Available scenarios

In this repository, we propose 5 intersection scenarios based on T/CMAX standard. All 5 scenarios are added to our test set, users can select desired scenarios form
./test/scenario_runner/srunner/examples/IntersectionContinuousTraffic.xml
Testing params are available for all 5 scenarios.

3 Intersection Scenarios

In this part, a series of scenarios of urban intersection are deployed to validate the performance of the RL agent. We develop traffic scenarios using CARLA scenario_runner module.

We select 5 challenging scenarios of intersection scenarios:

• turning left meeting going straight traffic flow

• turning left meeting turning right traffic flow

• turning right meeting going straight traffic flow

• going straight meeting going straight traffic flow

• going straight meeting turning left traffic flow

All these 5 intersection scenarios are facing great challenges from collision risk. The autonomous vehicle must find a proper timing to cross the traffic flow to finish the determined route.

4 Baselines & Results

In our work, we deploy a td3 agent to handle the intersection turning task. Here are the testing results of our proposed traffic scenarios.

Besides, we provide 2 rules-based method as baselines, which are AEB agent and IDM agent. You can find them in train/rl_agents/challenge_agents/baselines

Citation