Home

Awesome

COiLTRAiNE: Conditional Imitation Learning Training Framework

This repository can be used to easily train and manage the trainings of imitation learning networks jointly with evaluations on the CARLA simulator. Objectives:

System Overview

COIL Diagram

The COiLTRAiNE framework allows simultaneous training, driving on scenarios in CARLA and prediction of controls on some static dataset. This process can be done on several experiments at the same time.

Getting started

Requirements

Installation

To install COiLTRAiNE, we provide a conda environment requirements file. Start by cloning the repository on some folder and then, to install, just run:

conda env create -f requirements.yaml

Setting Environment/ Getting Data

The first thing you need to do is define the datasets folder. This is the folder that will contain your training and validation datasets

export COIL_DATASET_PATH=<Path to where your dataset folders are>

Download a sample dataset pack, with one training and two validations, by running

python3 tools/get_sample_datasets.py

The datasets; CoILTrain , CoILVal1 and CoILVal2; will be stored at the COIL_DATASET_PATH folder.

To collect other datasets please check the data collector repository. https://github.com/carla-simulator/data-collector

Getting CARLA

Note: the automatic scenario evaluation only works for CARLA 0.8.x, however you can train and evaluate agents in CARLA 0.9.X.

For doing scenario evaluation in CARLA you must install CARLA 0.8.4 or CARLA 0.8.2 under docker. This tutorial explains how to install CARLA under docker.

Executing

Assuming that you have CARLA docker with a docker image name as "carlasim/carla:version" , you can execute the coiltraine system by running:

python3 coiltraine.py --folder sample --gpus 0 -de TestT1_Town01 -vd CoILVal1 --docker carlasim/carla:version

Where the --folder sample is the experiment batch containing all the experiments that are going to be trained and validated. The TestT1 is a driving scenario on Town01, defined as one of the classes on the drive/suites folder. The validation datasets are passed as parameter with -vd and should be placed at the COIL_DATASET_PATH folder.

Expected outputs

You should see a colored screen on the terminal.

Initial

After finishing training and validation, the terminal screen should start driving look like as below.

Second

You will not see any CARLA server screen popping up since CARLA running under docker runs offscreen. Also note that this example trains on sample data and tests on a sample benchmark. Thus, the resulting driving model will be of poor quality. Please, test some of the models from the conditional models zoo to get high performance conditional imitation models.

Conditional Models Zoo