This repo is the implementation of the paper "Social behavior for autonomous vehicles". It contains NGSIM env that can replay vehicle trajectories in the NGSIM dataset while also simulate some interactive behaviors, as well as Maximum Entropy-based SVO prediction in this paper.

Install the dependent package

pip install -r requirements.txt

Download the NGSIM dataset from this website (export to csv file) and ngsim_data_process.py along with the path to the downloaded csv file (may take a while)

python ngsim_data_process.py --path [YOUR PATH]/Next_Generation_Simulation__NGSIM__Vehicle_Trajectories_and_Supporting_Data.csv --scene i-80

Choose some vehicle and predict their svo:

python main_ngsim_ma_predict_given_traj.py 

Some settings

• start_time：认为交互开始的那一步，一般轨迹都有1000~2000步

• duration：认为交互持续的时间

• save_dir

• scene：需要和data process的结果一致，如果使用us-101，则要重新data process

• veh_id：必须设置想观察的车的id，地图上只会有这些车。第一个元素即为ego car，用来以它为准对齐各个车的step。在目前给定轨迹的demo中区分ego的意义暂时不大。

• rollout step：对应论文中预测SVO用的一段sequence的长度

Examplar results are given in "docs/", by plotting the sampled NGSIM trajectories and SVO predictions like:

@article{DBLP:journals/pnas/SchwartingPAKR19,
  author    = {Wilko Schwarting and
               Alyssa Pierson and
               Javier Alonso{-}Mora and
               Sertac Karaman and
               Daniela Rus},
  title     = {Social behavior for autonomous vehicles},
  journal   = {Proc. Natl. Acad. Sci. {USA}},
  volume    = {116},
  number    = {50},
  pages     = {24972--24978},
  year      = {2019},
}

This repo is released under the MIT License. The NGSIM data processing code is borrowed from NGSIM interface. The NGSIM env is built on top of highway env which is released under the MIT license.