FormulaZero

This is the reference implementation for our paper:

FormulaZero: Distributionally Robust Online Adaptation via Offline Population Synthesis PDF

Aman Sinha*, Matthew O'Kelly*,Hongrui Zheng*, Rahul Mangharam, John Duchi, Russ Tedrake

Abstract: Balancing performance and safety is crucial to deploying autonomous vehicles in multi-agent environments. In particular, autonomous racing is a domain that penalizes safe but conservative policies, highlighting the need for robust, adaptive strategies. Current approaches either make simplifying assumptions about other agents or lack robust mechanisms for online adaptation. This work makes algorithmic contributions to both challenges. First, to generate a realistic, diverse set of opponents, we develop a novel method for self-play based on replica-exchange Markov chain Monte Carlo. Second, we propose a distributionally robust bandit optimization procedure that adaptively adjusts risk aversion relative to uncertainty in beliefs about opponents’ behaviors. We rigorously quantify the tradeoffs in performance and robustness when approximating these computations in real-time motion-planning, and we demonstrate our methods experimentally on autonomous vehicles that achieve scaled speeds comparable to Formula One racecars..

Citing

If you find this code useful in your work, please consider citing:

	@article{sinha2020formulazero,
	title={FormulaZero: Distributionally Robust Online Adaptation via Offline Population Synthesis},
	author={Sinha, Aman and O'Kelly, Matthew and Zheng, Hongrui and Mangharam, Rahul and Duchi, John and Tedrake, Russ},
	journal={arXiv preprint arXiv:2003.03900},
	year={2020}
	}

Dependencies

Requires:

• docker
• nvidia-docker2
• A recent Nvidia GPU e.g. GTX980 or better.
• Ubuntu

Installation

The docker image essentially packages all dependencies in a safe environment. The scripts we provide will externally mount our source code, and our models, into the docker environment.

Docker and Nvidia-Docker

The following is all of the steps to build a docker image for FormulaZero from a fresh Ubuntu installation:

1. Install Docker for Ubuntu. Make sure to sudo usermod -aG docker your-user and then not run below docker scripts as sudo
2. Install nvidia-docker. Make sure to use nvidia-docker2 not nvidia-docker1.

sudo apt-get install -y nvidia-docker2

You can test that your nvidia-docker installation is working by running

nvidia-docker run --rm nvidia/cuda nvidia-smi

If you get errors about nvidia-modprobe not being installed, install it by running

sudo apt-get install nvidia-modprobe

and then restart your machine.

Building the FormulaZero Docker Image

To run the demo, do the following. You will need to have docker installed and have an Nvidia GPU. Currently, this only works on Linux machines (in particular the visualization does not work on Mac).

• In your terminal, navigate to the Simulator folder.

• Run ./download_files.sh

• If this script does not work for some reason, you can manually download the files here: https://drive.google.com/drive/folders/1cBRKoQ31lGhFYXkCiBOdl34JZvHwBQh0 And then place them manually in the following locations:

  • Unzip flow_weights.zip and then place in /python
  • Place map1_speed.msgpack and map1_range.msgpack in /python
  • Place lut_inuse.npz in python/mpc

• Run ./build_docker_ui.sh Note you may need to use sudo with this command depending on the way you installed docker.

• If you have docker 19.03 or later, install Nvidia Container Toolkit following the instructions here: (https://github.com/NVIDIA/nvidia-docker), and run ./docker_ui.sh. Note that you may need to use sudo with this command depending on the way you installed docker.

• If you have an older version of docker and are using nvidia-docker 1.0, run ./docker_ui_nvidia_docker_1.0.sh. Note that you may need to use sudo with this command depending on the way you installed docker.

• If you have an older version of docker and are using nvidia-docker 2.0, run ./docker_ui_nvidia_docker_2.0.sh. Note that you may need to use sudo with this command depending on the way you installed docker.

The code will take up to 5 minutes to compile upon startup depending on your machine architecture.