Interpretable End-to-end Autonomous Driving

[Project webpage] [Paper]

This repo contains code for Interpretable End-to-end Urban Autonomous Driving with Latent Deep Reinforcement Learning. This work introduces an end-to-end autonomous driving approach which is able to handle complex urban scenarios, and at the same time generates a semantic birdeye mask interpreting how the learned agents reasons about the environment. This repo also provides implementation of popular model-free reinforcement learning algorithms (DQN, DDPG, TD3, SAC) on the urban autonomous driving problem in CARLA simulator. All of the algorithms take raw camera and lidar sensor inputs.

System Requirements

• Ubuntu 16.04
• NVIDIA GPU with CUDA 10. See GPU guide for TensorFlow.

Installation

1. Setup conda environment

$ conda create -n env_name python=3.6
$ conda activate env_name

2. Install the gym-carla wrapper following the installation steps 2-4 in https://github.com/cjy1992/gym-carla.

3. Clone this git repo to an appropriate folder

$ git clone https://github.com/cjy1992/interp-e2e-driving.git

4. Enter the root folder of this repo and install the packages:

$ pip install -r requirements.txt
$ pip install -e .

Usage

1. Enter the CARLA simulator folder and launch the CARLA server by:

$ ./CarlaUE4.sh -windowed -carla-port=2000

You can use Alt+F1 to get back your mouse control. Or you can run in non-display mode by:

$ DISPLAY= ./CarlaUE4.sh -opengl -carla-port=2000

It might take several seconds to finish launching the simulator.

2. Enter the root folder of this repo and run:

$ ./run_train_eval.sh

It will then connect to the CARLA simulator, collect exploration data, train and evaluate the agent. Parameters are stored in params.gin. Set train_eval.agent_name from ['latent_sac', 'dqn', 'ddpg', 'td3', 'sac'] to choose the reinforcement learning algorithm.

3. Run tensorboard --logdir logs and open http://localhost:6006 to view training and evaluation information.

Trouble Shootings

1. If out of system memory, change the parameter replay_buffer_capacity and initial_collect_steps the function tran_eval smaller.

2. If out of CUDA memory, set parameter model_batch_size or sequence_length of the function tran_eval smaller.

Citation

If you find this useful for your research, please use the following.

@article{chen2020interpretable,
  title={Interpretable End-to-end Urban Autonomous Driving with Latent Deep Reinforcement Learning},
  author={Chen, Jianyu and Li, Shengbo Eben and Tomizuka, Masayoshi},
  journal={arXiv preprint arXiv:2001.08726},
  year={2020}
}