CV

expand all

Carnegie Mellon University

PhD

Machine Learning Department

Carnegie Mellon University

Undergraduate

Bachelor of Science in Artificial Intelligence; additional major in Discrete Mathematics and Logic

GPA: 4.0/4.0

Understanding visual attention beehind bee-inspired UAV navigation

Bioengineering Group, University of New South Wales - PI: Sridhar Ravi

• Used the attention patterns of trained Reinforcement Learning (RL) agents to infer how a real bee makes movement decisions
• Built a goal-conditioned RL environment in OpenAI Gym to train a UAV to imitate bee behaviors using bee-like input sensors
• Used SHAP values, a tool for explaining model output, to measure visual regions that trained RL agents pay attention to

AlephZero: Extending AlphaZero to Infinite Boards

Independent Research - PI: Pranav Rajbhandari

• Defined and analyzed $\aleph_0$ board games, a class of games with potentially unbounded action spaces. Interesting examples include 'Jenga' and '5D Chess with Multiverse Time Travel', as well as classic games like 'Chess' and 'Tic-Tac-Toe'
• Developed AlephZero, an extension of AlphaZero able to learn optimal policies in $\aleph_0$ board games
• Utilized transformer architectures to define policy networks and value networks able to take multi-dimensional sequential input
• Compared approach to standard algorithms such as AlphaZero, Deep Q-Learning, and Monte Carlo Tree Search

Fine Tuning Swimming Locomotion Learned from Mosquito Larvae

University of New South Wales; U.S. Naval Research Laboratory - PI: Sridhar Ravi; Donald Sofge

• Optimized swimming locomotion copied from mosquito larvae for use on a robotic platform
• Utilized Reinforcement Learning to guide a local search algorithm optimizing swimming locomotion
• Designed an OpenAI Gym environment utilizing a Computational Fluid Dynamics (CFD) model for training
• Sped up the training process by using a pre-trained deep neural network to accurately predict forces on a robotic swimmer
• Compared performance of various architectures, including Deep Neural Networks, Recurrent Neural Networks, and LSTMs

Transformer guided coevolution: Team selection in multiagent adversarial games

U.S. Naval Research Laboratory - PI: Donald Sofge; Prithviraj Dasgupta

• Developed BERTeam, an algorithm to learn diverse and cooperative team selection for multiagent adversarial team games
• Evaluated algorithm on Pyquaticus, a simulation of robotic Marine Capture-The-Flag
• Used Masked Language Modeling to teach optimal team composition to BERTeam's transformer architecture
• Cotrained BERTeam with Coevolutionary Deep Reinforcement Learning to select teams from a diverse population of agents
• Compared result of training with established algorithms in literature
• Developed and maintained unstable_baselines3, a Python package extending stable_baselines3 to multiagent environments

Geodesic complexity? It's actually quite simplex

Department of Mathematical Sciences, Carnegie Mellon University - PI: Florian Frick

• Explored geodesic complexity, a measure of difficulty for creating an efficient continuous motion plan on a metric space
• Designed a technique utilizing local properties of a space to lower bound its geodesic complexity
• Created and proved correctness of an algorithm calculating cut loci on surfaces of polyhedra, a property related to their geodesic complexity
• Applied these techniques to produce a novel result for the geodesic complexity of the octahedron
• Proved existing geodesic complexity bounds in a new way, displaying the utility of our general method

Utilizing Sim-to-Real Methods for Training a Robot Arm

Reliable Autonomous Systems Laboratory, Carnegie Mellon University - PI: Reid Simmons

• Led a team of four to design and maintain an OpenAI Gym environment for a Kinova Jaco Gen3 6DOF robot arm
• Simulated a model of the robot arm compatible with the control scheme of the physical arm using the Gazebo simulator
• Utilized ROS to handle communication between the robot arm and Python scripts
• Trained a 'real life filter' with the CycleGAN algorithm to make photo-realistic simulation images used for training
• Implemented a training pipeline for a robotic manipulation task, trained in simulation and refined on the real arm

Learning NEAT Emergent Behavior in Robot Swarms

Distributed Autonomous Systems Group, U.S. Naval Research Laboratory - PI: Donald Sofge

• Developed an algorithm for training local policies to produce emergent behaviors in a robot swarm
• Designed a training pipeline applying the NeuroEvolution of Augmenting Topologies (NEAT) algorithm to robot swarm control
• Tested the algorithm's performance on a variety of tasks and simulated robotic swarms using the CoppeliaSim simulator
• Utilized ROS to handle communication between Python scripts and robotic swarms (both real and simulated)

UAV Routing for Enhancing the Performance of a Classifier-in-the-loop

Distributed Autonomous Systems Group, U.S. Naval Research Laboratory - PI: Swaroop Darbha

• Collaborated on an interdisciplinary research project optimizing the information gained from targets by robot swarms
• Designed a heuristic algorithm for planning robot paths inspired by approximate solutions to the Traveling Salesman Problem
• Utilized Mathematica software, as well as methods from 'Convex Optimization' to optimize solutions for large test cases
• Tested our algorithm on both generated and real-life problem instances using Julia and the Gurobi optimizer

Comparing Transfer Learning Methods for Continuous Reinforcement Learning

Adaptive Systems Section, U.S. Naval Research Laboratory - PI: Laura Hiatt

• Planned and executed a research project evaluating various transfer learning methods on robot arm manipulation tasks
• Designed an OpenAI Gym environment for a robotic manipulation task using the MuJoCo simulator
• Compared the performance of known transfer learning methods in transferring knowledge between Deep Neural Networks
• Utilized ROS to handle communication between the robot arm and Python scripts

Creating a Strategic Agent to Play Jenga

Reliable Autonomous Systems Laboratory, Carnegie Mellon University - PI: Reid Simmons

• Planned and executed a research project evaluating the performance of various adversarial AI algorithms playing Jenga
• Implemented algorithms such as Monte Carlo Tree Search, Deep Q-Networks, and Inverse Reinforcement Learning
• Created a statistical model to estimate the stability of a Jenga tower for use in Model Based Reinforcement Learning
• Trained the model through repeatedly sampling stabilities of towers with the PyBullet physics engine

Geodesic complexity of the octahedron, and an algorithm for cut loci on convex polyhedra

Florian Frick and Pranav Rajbhandari

arXiv

(Preprint, submitted to Journal of Applied and Computational Topology)

Understanding visual attention beehind bee-inspired UAV navigation

Pranav Rajbhandari, Abhi Veda, and 3 more authors

arXiv

(Preprint, poster accepted to 2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2025))

Transformer Guided Coevolution: Improved Team Formation in Multiagent Adversarial Games

Pranav Rajbhandari, Prithviraj Dasgupta, and Donald Sofge

Proc. of the 24th International Conference on Autonomous Agents and Multiagent Systems (AAMAS 2025)

Fine Tuning Swimming Locomotion Learned from Mosquito Larvae

Pranav Rajbhandari, Karthick Dhileep, and 2 more authors

2024 IEEE International Conference on Robotics and Biomimetics (ROBIO)

Learning NEAT Emergent Behaviors in Robot Swarms

Pranav Rajbhandari and Donald Sofge

2024 IEEE International Conference on Robotics and Biomimetics (ROBIO)

UAV Routing for Enhancing the Performance of a Classifier-in-the-loop

Deepak Prakash Kumar, Pranav Rajbhandari, and 3 more authors

2024 IEEE International Conference on Robotics and Biomimetics (ROBIO)

Learning Emergent Behavior in Robot Swarms with NEAT

Pranav Rajbhandari and Donald Sofge

Naval Applications of Machine Learning

Sim-to-real Transfer Reinforcement Learning

Pranav Rajbhandari, Sophia Zalewski, and Reid Simmons

Carnegie Mellon University Meeting of the Minds

Creating Agents to Learn Jenga

Pranav Rajbhandari and Reid Simmons

Carnegie Mellon University Meeting of the Minds

Researcher

University of New South Wales - Bioengineering Group

• Worked in the Bioengineering Group on project "Understanding visual attention beehind bee-inspired UAV navigation"
• Worked in the Bioengineering Group on project "Fine Tuning Swimming Locomotion Learned from Mosquito Larvae"

Researcher

U.S. Naval Research Laboratory

• Worked in the Distributed Autonomous Systems Group on project "Transformer guided coevolution: Team selection in multiagent adversarial games"
• Worked with USNW Canberra on project "Fine Tuning Swimming Locomotion Learned from Mosquito Larvae"
• Worked in the Distributed Autonomous Systems Group on project "Learning NEAT Emergent Behavior in Robot Swarms"
• Worked with Texas A&M Department of Mechanical Engineering on project "UAV Routing for Enhancing the Performance of a Classifier-in-the-loop"
• Worked in the Adaptive Systems Section on project "Comparing Transfer Learning Methods for Continuous Reinforcement Learning"

Researcher

National Aeronautics and Space Administration - Ames Research Center

• Created an AI system to automate calling airport TMI events, especially Ground Stops and Ground Delay Programs
• Explored Imitation Reinforcement Learning methods to compete against the baseline of training a classifier model
• Processed historical data and created models to approximate decision processes using Python and R

Researcher

Carnegie Mellon University

• Worked in Department of Mathematical Sciences on project "Geodesic complexity? It's actually quite simplex"
• Worked in the Reliable Autonomous Systems Laboratory on project "Utilizing Sim-to-Real Methods for Training a Robot Arm"
• Worked in the Reliable Autonomous Systems Laboratory on project "Creating a Strategic Agent to Play Jenga"

Teaching Assistant

Carnegie Mellon University

For 'AI: Representation and Problem Solving' (3 semesters), 'Concepts of Mathematics' (1 semester), and 'Probability Theory for Computer Scientists' (1 semester)

• Collaborated in a team of up to 10 Teaching Assistants to manage classes of up to 100 students
• Planned and led class-wide review sessions, as well as recitations of about 20 students
• Held office hours to help students understand course material in a one-on-one setting
• Created, tested, and graded programming assignments and written homework

Research Assistant

Carnegie Mellon University

• Collaborated with a team of three researchers to develop and maintain an R package for Natural Language Processing
• Utilized Rust's BERT Natural Language Processing to tokenize and classify strings in R

Programmer

Centers for Disease Control and Prevention - Chronic Viral Diseases Branch Immunology Lab

• Designed a Constraint Satisfaction Problem instance to automate generating laboratory experiment setup procedures
• Utilized Python and R to automate post-experiment data processing
• Refined and deployed these programs across the laboratory after prototyping and incorporating feedback from lab members

Dean’s List, High Honors (8 semesters)

Carnegie Mellon University

Senior Leadership Recognition Award

Carnegie Mellon University

Dr. William Brown Academic Achievement Award

Carnegie Mellon University

Tartan Leaders of Tomorrow

Carnegie Mellon University

Winner of AI/ML Innovation Challenge

Naval Surface Warfare Center Dahlgren Division

Was awarded \$50,000 cash prize at three-day competition hosted by the US Navy; Designed algorithm to protect ships from enemy missiles

Carnegie Mellon University Super Informal Topology Discussion Group

Presenter, Member

Carnegie Mellon University Track & Field

Sprint Team Captian

Carnegie Mellon University PRISM Club

Volunteer, Member