Skip to content

Latest commit

 

History

History
executable file
·
123 lines (79 loc) · 3.79 KB

README.md

File metadata and controls

executable file
·
123 lines (79 loc) · 3.79 KB
Raw

InfoGAIL: Interpretable Imitation Learning from Visual Demonstrations

By Yunzhu Li, Jiaming Song, Stefano Ermon

Introduction

Modified codebase of TORCS, with the ability to extract dashboard views.

InfoGAIL implementation, attached with two examples: pass & turn.

Citing InfoGAIL

If you find this codebase useful in your research, please consider citing:

@article{li2017inferring,
    title={InfoGAIL: Interpretable Imitation Learning from Visual Demonstrations},
    author={Li, Yunzhu and Song, Jiaming and Ermon, Stefano},
    journal={arXiv preprint arXiv:1703.08840},
    year={2017}
}

Requirements

  1. Python 2.7
  2. Tensorflow 0.12.1
  3. Keras 1.2.2
  4. xautomation

Install and Configure TORCS

In the following section, we only show the commands for running example 0 (indicating pass). Users can replace 0 with 1 to run example 1 (indicating turn).

  1. Install all the dependencies of TORCS, including but not limited to:
sudo apt-get install libplib-dev
sudo apt-get install libopenal-dev
sudo apt-get install libalut-dev

See http://torcs.sourceforge.net/ for more information

  1. cd to the torcs-1.3.4/ folder, type the following commands to install TORCS:
./configure
make
sudo make install
sudo make datainstall

Default installation directories are:

/usr/local/bin
/usr/local/lib/torcs
/usr/local/share/games/torcs

Run the torcs command to play TORCS.

  1. Copy the modified tracks files in the modified_tracks/ folder to the torcs folder /usr/local/share/games/torcs/tracks/road
  2. Type the following commands to configure the running environment:
rm -rf ~/.torcs
cp -r torcs_config_ai_0 ~/.torcs

Download training data and pretrained weights

  1. Training data
  1. Pretrained weights

Run with pretrained weights

  1. cd to wgail_info_0/

  2. open drive.py and edit line 20-21 in concert with downloaded data and weights

  3. change variable code in line 14 into different values (0 or 1) to observe different behaviors

  4. type the python drive.py to run pass with pretrained weights

    Run pass with different latent codes (0 or 1):

    Run turn with different latent codes (0 or 1):

Training

  1. cd to wgail_info_0/
  2. open wgail_info.py and edit line 17-19 in concert with downloaded data and weights
  3. open models.py and edit line 508 to specify the place to store the weights, edit line 422 to specify the place to store the log file
  4. type python wgail_info.py to train on pass, note that the weights trained via behavior cloning are used to initialize the policy network

Misc.

  1. Track selection:

    • pass: chenyi-Street 1
    • turn: chenyi-Brondehach

  2. torcs-1.3.4 in this repository is a modified version of the original codebase released by Bernhard Wymann. This version holds the ability of extracting and transmitting visual information.

References

DeepDriving

Using Keras and Deep Deterministic Policy Gradient to play TORCS

TORCS - The Open Race Car Simulator

SnakeOil

Simulated Car Racing Championship Competition Software Manual