Paper

https://arxiv.org/abs/2410.10646

Video

https://youtu.be/GUZlGBk60uY

Code Setup Instructions

Assuming your current directory path is 'clone_location'.

1. Create and activate the conda environment
   Ensure you have conda installed. Then, run the following commands to create and activate the environment:

   git clone https://github.saobby.my.eu.orge/James-R-Han/DR-MPC
   cd clone_location/DR-MPC
   conda env create -f environment.yml
   conda activate social_navigation

   Note: you may need to adjust the torch and cuda version to match your device

2. Clone and install RVO2
   Next, clone the Python RVO2 repository and install it:

   Note: you should have cmake installed.

   sudo apt update
   sudo apt install cmake
   

   git clone https://github.com/sybrenstuvel/Python-RVO2.git
   cd Python-RVO2
   python setup.py build
   python setup.py install

3. Clone and install pysteam
   Clone the pysteam repository and install it:

   cd clone_location/DR-MPC
   git clone https://github.com/utiasASRL/pysteam.git

4. Setup PYTHONPATH
   Navigate to the DR-MPC directory and update your PYTHONPATH:

   cd clone_location/DR-MPC
   export PYTHONPATH="${PYTHONPATH}:clone_location/DR-MPC"

5. Run the training script
   This command will train a policy specified in 'scripts/configs/config_general' for a number of trials on different seeds specified in 'scripts/configs/config_training'

   Note: for video generation you'll need ffmpeg

   sudo apt install ffmpeg

   cd clone_location/DR-MPC
   python3 scripts/online_continuous_task.py

6. Comparing the results of different training runs An example is already setup in scripts/compare_training_multirun.py

   python3 scripts/compare_training_multirun.py

The codebase is largely seprated into the environment ('environment') and training ('scripts'). The environment is modularized into the path tracking and human avoidance components as indicated by the folder names. RL learning algo (SAC), models (DRL, ResidualDRL, DR-MPC), OOD pipeline, etc. are all contained in the 'scripts' folder.