Artificial Intelligence & Machine Learning Research Project - Comparison of Vision Models for Traversability Estimation
This repository contains the code and experiments for the research project:
"Comparison of Specialized and Prompt-Driven Vision Models for Off-Road Traversability Estimation on RELLIS-3D"
Conducted as part of the COMP SCI 7205: Artificial Intelligence & Machine Learning Research Project course in the Master of Artificial Intelligence and Machine Learning program at the University of Adelaide.
Autonomous robots operating in off-road environments must determine which terrain regions are safe to traverse. This task is known as traversability estimation.
This project investigates vision-based methods for terrain understanding, evaluating approaches that predict whether terrain is traversable using RGB imagery.
The project focuses on:
- Evaluating computer vision techniques for terrain segmentation
- Analyzing qualitative and quantitative traversability predictions
- Studying failure cases and robustness in real-world environments
- Investigating datasets used for off-road robot navigation research
The project contributes both experimental outputs and a comprehensive research report summarizing methodology, evaluation, and findings.
.
├── ckpts/ # Model checkpoints used during experiments
├── outputs/
│ └── prediction/ # Generated qualitative predictions
├── rellis-output/ # Output results from RELLIS-3D dataset experiments
├── results/
│ └── plots/ # Evaluation plots and visualizations
├── scripts/ # Utility scripts for running experiments
│
├── Analysis Presentation.pdf # Project presentation slides
├── Final Report.pdf # Final research report
├── Literature Review.pdf # Literature review document
├── README.md # Project README
└── Research Proposal.pdf # Initial research proposal
- Clone this repository:
git clone https://github.com/BrownAssassin/AIML-Research-Project.git
cd AIML-Research-Project- Create a virtual environment (recommended):
conda create -n traversability python=3.10
conda activate traversability
- Install dependencies:
pip install numpy pandas matplotlib opencv-python torch torchvision tqdm*Additional packages may be required depending on the scripts used.
Experiments in this project require the use of off-road robotics datasets designed for terrain perception research.
A multi-modal dataset for off-road autonomous navigation containing RGB imagery, LiDAR, and semantic annotations.
Source: GitHub
Dataset outputs generated during experimentation are stored in:
rellis-output/
Scripts for generating predictions and evaluation plots are located in the scripts/ directory.
Typical workflow:
- Load dataset with
download_rellis3d_from_readme.sh - Verify dataset integrity with
rellis3d_sanity_check.py - Run GSAM inference pipeline script (
rellis3d_gsam_singleprocess.py) - Generate predictions and plots (
rellis3d_list_ids.py,rellis3d_eval_binary.py,plot_metrics.py)
This repository includes all major deliverables produced during the research project:
- Literature Review
- Research Proposal
- Analysis Presentation
- Final Research Report
These documents provide a detailed explanation of:
- Prior work in traversability estimation
- Experimental methodology
- Model evaluation
- Discussion of results and future research directions
This repository is intended for academic and educational use.
If datasets or third-party models are used, please refer to their respective licenses.
Mrinank Sivakumar
Master of Artificial Intelligence and Machine Learning, University of Adelaide
Bachelor of Information Technology (Honours), University of Ontario Institute of Technology