Social Driving

Design multi-agent environments and simple reward functions such that social driving behavior emerges

Snapshot of 7 agents in a nuScenes Intersection

Installation

Create a new Conda environment

$ conda create --name sdriving python=3.8
$ conda activate sdriving

Follow the instructions given here to install PyTorch. Install PyTorch version 1.5.2 or 1.7 if you want to use JIT compilation which speeds up the code by roughly 1.5-2x.

Install the Nuscenes DevKit for access to realistic simulation environments. Use the master branch instead of the pip version. Additionally, download the map expansion pack.

Finally, install sdriving using the following instructions.

$ git clone https://github.com/fidler-lab/social-driving.git sdriving
$ cd sdriving
$ python setup.py develop

If you need to train an agent, some additional configuration is needed. You can skip these steps if you are only interested in generating rollouts / don't care about logging on the server (using WANDB_MODE=dryrun).

1. Create a Wandb account using wandb login and following the steps given there.

2. A project named social-driving will be automatically created. If anything goes wrong here please open an issue.