Skip to content

Latest commit

 

History

History
84 lines (68 loc) · 3.69 KB

README.md

File metadata and controls

84 lines (68 loc) · 3.69 KB
Raw

DistributionallyRobust.jl

This is official code repository of "Distributionally Robust Risk-Aware Control Framework for Safe Crowd Navigation with Human Motion Predictions".
This code is based on code of "Risk Sensitive Sequential Action Control" (MIT License) RSSAC code.

DRCC-MPC example in eth dataset DRCC-MPC example in hotel dataset

Installation

Tested on Julia 1.7.3 & Python 3.6
This is summary for installing all submodules and dependencies.

Cloning

Clone the submodules when cloning this repository. Submodules included in this repository is

  • Trajectron-plus-plus
  • CrowdNav
  • Python-RVO2
git clone --recurse-submodules <repository cloning URL>

Environment Setup

Since Trajectron-plus-plus module is based on python 3.6, which is bit outdated, we recommend to use conda environment for environment setup.
First, create conda environment and install dependencies for Trajectron++.

conda create --name [your conda environment name] python=3.6 -y
conda activate [your conda environment name]
cd Trajectron-plus-plus
pip install -r requirements.txt

We have to install Python-RVO2 library to install CrowdNav.
We can start with installing CMake and Cython.
Install the tested version of Cython for Python-RVO2.

cd ../Python-RVO2
pip install -r requirements.txt

Build and install.

python setup.py build
python setup.py install

If Python-RVO2 is installed successfully, we can install CrowdNav.

cd ../CrowdNav
pip install -e .

Submodule Training

Trajectron++

We have to get trained Trajectron++ module for our DRCC-MPC.
Training procedure for Trajectron++ is givn in its repository.
We are only working with Pedestrian Dataset in here.

Load trained Trajectron++ module

You can access to pretrained Trajectron++ module in RSSAC repository.
They are under Trajectron-plus-plus/experiments/pedestrians/models. Download trained model and put it in same directory to use the model.

CrowdNav

You can also follow the CrowdNav training procedure.
However, it is not necessary if you are not planning to to experiments with CrowdNav. You may have to change training configuration in crowd_nav/configs.
For example, if you want to test CrowdNav with pedestrian dataset, you should change time_step = 0.4 in env.config to comply with dataset timestep.

Install Julia

We can install Julia from official webpage.
Note that this module is tested with Julia 1.7.3
You may want to install julia within your conda environment but it is not necessary.

Julia can call python function using PyCall. Our integration with Trajectron++ and CrowdNav is achieved through PyCall.
If you open notebook file in editor after activating conda environment, it will automatically detect python version in conda and use it for PyCall.

Compatability issue

Plot.jl now doesn't support matplotlib < 3.4.0
Since python 3.6 does not support matplotlib >= 3.4.0, we had to go around this compatability issue.
One temporary solution is save result data in corrent conda environment and plot in other conda environment with newer matplotlib.

  1. Save prediction and robot history data to csv file
  2. Open python_plot.ipynb to load the data in python and make gif