Hi, I'm Aravindhan and I love to make Machines Learn!

Machine Learning Engineer with research experience in computer vision and reinforcement learning, currently pursuing MS in Robotics and Autonomous Systems while working as a Software Engineer in Research Development at Arizona State University. Passionate about developing robust AI systems for real-world applications, with a focus on multi-agent systems, visual tracking, and autonomous navigation.

Research Projects

Q-Learning with LLMs (Ongoing) | Sep '24 - Present

Tech Stack: Python, Q-Learning, LLAMA, LORA , OpenAI Gym

Developing an innovative RL model using the LLAMA language model to replace the Bellman operator for generating Q-tables in reinforcement learning tasks
Implementing and testing RL agents across four OpenAI Gym environments (Lunar Lander, Inverted Pendulum, Mountain Car, CartPole) to evaluate model performance and convergence
Currently integrating LoRA (Low-Rank Adaptation) for efficient fine-tuning of the LLAMA model, optimizing it specifically for RL tasks with minimal computational overhead
Demonstrating improved sample efficiency and faster convergence compared to traditional Q-learning approaches across multiple test environments

BeMaD - Behavioral Malware Detection

Tech Stack: Python, Q-Learning, GANs, RQN (Recurrent Q-Networks), GRU, PostgreSQL

Engineered a novel malware detection system combining Q-learning with GANs, achieving 97% accuracy on the JUGAAD dataset through a unique three-way update mechanism (Q-network, Generator, Discriminator) for robust malware variant detection.
Implemented a hierarchical classification model using RQNs with GRU layers (100 units x 4) and supervised learning, resulting in 94.71% F1-score on CICAndMal2017 dataset while maintaining a lightweight 1.2MB model size.
Designed a platform-agnostic architecture supporting both Windows and Android environments, processing network traffic patterns in real-time through dense feature extraction layers and Q-learning style parameter updates (θt+1 = θt + α · ∇θL(θt)).
View paper here

Safe Multi-Robot Navigation with Deep RL

Tech Stack: Python, OpenAI Gym, PyTorch, Stable Baselines3, ROS, Control Barrier Functions, Robotarium

Implemented a collision avoidance system integrating deep reinforcement learning with control barrier functions for multi-robot navigation, achieving 93% success rate in challenging high-density environments.
Built a custom OpenAI Gym environment for simulating aerial robots, and conducted extensive experiments with various RL algorithms (PPO2, A2C, DQN, TRPO) across 1000+ evaluation episodes to benchmark performance.
Enhanced navigation performance using PPO2, reducing collision rates from 12% to 2.8% while improving path efficiency by 25%, validated through systematic testing in simulated aerial robot scenarios with 40% spatial density.

Past Projects

Robust Visual Tracking Through Atmospheric Turbulence (Ongoing) Tech Stack: PyTorch, DINO-Tracker, WGAN-GP, OpenCV, Python

Currently enhancing the Restormer architecture by incorporating WGAN-GP and multi-scale processing to improve tracking performance under turbulent conditions, where traditional methods show significant accuracy drops.
Developing a benchmark dataset for evaluating point tracking in turbulent scenarios by applying turbulence simulations to the TAP-Vid dataset, aiming to enable systematic testing across varying intensities.
Implementing turbulence-specific adaptations to the DINO-Tracker framework, with preliminary results showing improvements in position accuracy (72.3%) and occlusion handling (65.8%) under moderate turbulence conditions.

Two Player Soccer - Multi-Agent RL Environment

Tech Stack: Python, PyTorch, OpenAI Gym, Stable Baselines3, Pygame

Designed and implemented a zero-sum competitive environment for multi-agent reinforcement learning, featuring custom reward mechanisms for ball possession, scoring, and agent interactions.
Created an OpenAI Gym-compatible environment with discrete action spaces and custom observation states, enabling systematic training and evaluation of various RL algorithms for competitive gameplay.
Developed stochastic initialization protocols and penalty systems to ensure balanced gameplay, incorporating collision detection and ball transfer mechanics for realistic soccer-like interactions between agents.

Medical Image Representation Learning with SimCLR Tech Stack: PyTorch, ResNet-18, NIH Chest X-ray Dataset, Python

Implemented the SimCLR framework for self-supervised medical image representation learning, utilizing ResNet-18 backbone and Gaussian blurring augmentation on the NIH chest X-ray dataset.
Developed custom training pipeline incorporating contrastive loss and Euclidean distance metrics, achieving 93.6% training accuracy and 61% validation accuracy on representation quality evaluation.
Conducted systematic experiments analyzing the impact of various data augmentation strategies and loss function configurations on model performance and representation quality.

Publications

Aravindhan Thaninayagam, D.O, Prasath, S.V.R, & Roshan. (2022). Collision Avoidance System using Reinforcement Learning. 4th International Conference on Communication and Intelligent Systems (ICCIS).

Get in touch

Email : Aravindhan@asu.edu Call: +1 602-566-1854

Page updated

Google Sites

Report abuse