OATomobile: A research framework for autonomous driving

Overview | Installation | Baselines | Paper

OATomobile is a library for autonomous driving research. OATomobile strives to expose simple, efficient, well-tuned and readable agents, that serve both as reference implementations of popular algorithms and as strong baselines, while still providing enough flexibility to do novel research.

Overview

If you just want to get started using OATomobile quickly, the first thing to know about the framework is that we wrap CARLA towns and scenarios in OpenAI gyms:

import oatomobile

# Initializes a CARLA environment.
environment = oatomobile.envs.CARLAEnv(town="Town01")

# Makes an initial observation.
observation = environment.reset()
done = False

while not done:
  # Selects a random action.
  action = environment.action_space.sample()
  observation, reward, done, info = environment.step(action)

  # Renders interactive display.
  environment.render(mode="human")

# Book-keeping: closes
environment.close()

Baselines can also be used out-of-the-box:

# Rule-based agents.
import oatomobile.baselines.rulebased

agent = oatomobile.baselines.rulebased.AutopilotAgent(environment)
action = agent.act(observation)

# Imitation-learners.
import torch
import oatomobile.baselines.torch

models = [oatomobile.baselines.torch.ImitativeModel() for _ in range(4)]
ckpts = ... # Paths to the model checkpoints.
for model, ckpt in zip(models, ckpts):
  model.load_state_dict(torch.load(ckpt))
agent = oatomobile.baselines.torch.RIPAgent(
  environment=environment,
  models=models,
  algorithm="WCM",
)
action = agent.act(observation)

Installation

We have tested OATomobile on Python 3.5.

1. install python

   # install python
   wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
   chmod 777 Miniconda3-latest-Linux-x86_64.sh
   ./Miniconda3-latest-Linux-x86_64.sh
   source ~/.bashrc
   
   conda create -n py35 python=3.5
   conda activate py35

2. To install the core libraries (including CARLA, the backend simulator):

   # The path to download CARLA 0.9.6.
   export CARLA_ROOT=...
   mkdir -p $CARLA_ROOT
   
   # Downloads hosted binaries.
   wget http://carla-assets-internal.s3.amazonaws.com/Releases/Linux/CARLA_0.9.6.tar.gz
   
   # CARLA 0.9.6 installation.
   tar -xvzf CARLA_0.9.6.tar.gz -C $CARLA_ROOT
   
   # Installs CARLA 0.9.6 Python API.
   easy_install $CARLA_ROOT/PythonAPI/carla/dist/carla-0.9.6-py3.5-linux-x86_64.egg
   
   # install dependencies
   pip install -r $PWD/oatomobile/requirements.txt
   pip install -r $PWD/oatomobile/requirements2.txt

3. To install the OATomobile core API:

   pip install --upgrade pip setuptools
   pip install oatomobile

4. To install dependencies for our PyTorch- or TensorFlow-based agents:

   pip install oatomobile[torch]
   # and/or
   pip install oatomobile[tf]

Citing OATomobile

If you use OATomobile in your work, please cite the accompanying technical report:

@inproceedings{filos2020can,
    title={Can Autonomous Vehicles Identify, Recover From, and Adapt to Distribution Shifts?},
    author={Filos, Angelos and
            Tigas, Panagiotis and
            McAllister, Rowan and
            Rhinehart, Nicholas and
            Levine, Sergey and
            Gal, Yarin},
    booktitle={International Conference on Machine Learning (ICML)},
    year={2020}
}

Modifications to OATomobile

I used OATomobile to generate data for my bachelor thesis and also modified it for that purpose. The scripts I wrote are very basic and not well designed. See generate_data.py for data generation. See train.py for training of a DIM model. A model can be trained as follows:

python train.py --dataset_dir=data/mydata/processed --output_dir=models/dim/mymodel_d32/ --num_epochs=201 --cuda_train_idx=0 --cuda_val_idx=0 --mobilenet_num_classes=32 --batch_size=512

. See evaluation.py for the evaluation of a trained DIM model. I also modified files of OATomobile such as oatomobile.datasets.carla.py.