/NGSIM_SMARTS

Primary LanguagePython

How to run

  1. Install SMARTS following the instructions in ./SMARTS folder

  2. Download the original data from https://jbox.sjtu.edu.cn/v/link/view/75a0931a222347e1ba2e0441407f4a1f and place it under ./ngsim folder

  3. Build NGSIM scenario with scl build --clean ./ngsim

  4. Generate expert demonstrations with python example_expert_generation.py

  5. Test rollout with python example_rollout.py

项目说明

项目实现

我们实现了使用GAIL+TD3算法来实现驾驶决策任务，同时并使用Behavior Clone算法先预训练TD3模型，使得生成出来的决策分布和专家分布相近，减缓分布差距过大导致的梯度消失问题。此外还引入梯度惩罚和梯度裁剪来使得训练更加稳定。

效果展示

快车道效果

images

慢车道效果

images