demo_enhanced_morl_nav

Code repository for the publication "Demonstration-Enhanced Adaptable Multi-Objective Robot Navigation" by Jorge de Heuvel, Tharun Sethuraman, and Maren Bennewitz, in Proceesings of IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2025.

Paper website: https://www.hrl.uni-bonn.de/publications/2025/deheuvel2025iros_morl

Setup

Prerequisites

In order to use the script setup.sh, Conda must be installed.

Steps

Simply create a new conda environment and install all dependencies by running:

bash setup.sh

Setup on Mac OS

During the time of code upload, a compiler issue due to an updated clang version on MacOS prevented installation of igibson and pybullet. The issue is described here: bulletphysics/bullet3#4712 If available, use a Linux system.

Instructions

There are multiple steps involved for training the approach:

1. Generate a demonstration dataset for the DREX behavior cloning (BC) policy.
2. Training of the DREX BC policy.
3. Generating a noise-infused dataset for the DREX reward-model by rolling out the DREX policy.
4. Training of the DREX reward model.
5. Preparing the PD-MORL preference space interpolator by training a set of preference key points.
6. Training the final PD-MORL policy.

All steps are summarized in the shell script training_pipeline_full.sh.