Traditional reinforcement learning often requires millions of interactions to learn effective control policies. This project investigates analytical policy gradients (APGs), which leverage differentiable physics to compute gradients directly through the dynamics, enabling faster and more sample-efficient policy learning. To support this research, I developed a GPU-accelerated differentiable quadrotor simulator in PyTorch capable of simulating 100,000+ drones in parallel, providing a fast and scalable platform for developing and evaluating autonomous flight algorithms.

The Crazyflie is a lightweight and inexpensive quadrotor that provides a safe platform for developing and testing new autonomy algorithms. I maintain our lab's Crazyflie platform, integrating the Crazyswarm software stack with our Vicon motion capture system and extending the hardware with a breakout deck and onboard third-person camera. Together, these capabilities provide a flexible testbed for experimenting with autonomous flight on resource-constrained hardware.

Crazyswarm flight with Vicon feedback

Two-drone figure-eight tracking

Live onboard camera feed streamed to a desktop

I spent the summer in Penn's ScalAR Lab developing hardware and autonomy for a full-scale Clearpath Heron autonomous surface vehicle (ASV). My work ranged from designing custom electronics for the robot to developing waypoint planning software for autonomous river surveys on the Schuylkill River.

Hardware: RS232 → USB PCB

I designed, fabricated, and assembled a custom PCB that converts RS232 to USB, allowing modern peripherals to interface with the Heron.

Schematic

PCB layout

Fabricated board

Autonomous Field Deployment

I also developed a lawnmower waypoint generator within the robot's ROS autonomy stack, enabling the Heron to autonomously survey bounded regions and collect depth measurements for riverbed reconstruction.

``` ```

Field Experiments, riverbed reconstruction, and waypoint planning.

I spent the summer at MIT's Space Enabled Media Lab working on Zero Robotics, an international programming competition that makes space robotics more accessible by giving students the opportunity to write code that is executed on the Astrobee robots aboard the International Space Station during the final competition. I helped design the competition challenges, tuned PID controllers for the Astrobee simulator that students programmed against, and helped host the 2022 ISS finals alongside NASA astronauts.

I led a small intern team on a search-and-rescue drone swarming project, using gym-pybullet-drones to prototype and evaluate A*-based multi-agent search in simulation. The goal was to test whether the simulator could serve as a viable platform for multi-agent robotics research.

Multi-drone swarm simulation in gym-pybullet-drones

Summer soldering practice project

I was contracted by a 10-million-subscriber YouTube creator to develop a fast-paced multiplayer FPS in Roblox Studio. I initially built the project as the sole developer, implementing the game's core systems, including weapons, combat mechanics, visual effects, HUDs, progression, and shop/inventory systems. As development progressed, I expanded the team to include additional contributors for content creation while continuing to lead development. The game reached over 500,000 plays in under two months and was nominated for Best Upcoming Experience at the Roblox Developers Conference (RDC).

Core FPS gameplay.

Candy Land map.

Blimp arena.

Main lobby.

To better understand how modern deep learning frameworks work under the hood, I built a neural network framework from scratch in Python. Rather than relying on automatic differentiation, I derived and implemented the analytical gradients for each layer by hand, implementing the forward and backward passes, optimization algorithms, and training loop from the ground up. This project laid the groundwork for a more general automatic differentiation framework that I'm currently developing.

Learned decision boundary on a nonlinear 5-class dataset.

Spiral classification — predicted / actual / errors.

MNIST confusion matrix after training.

Resources: Omar Aflak — Neural Network from Scratch in Python . This article was my primary reference while implementing the framework.

This project generates an interactive, physics-based 3D scene from a single RGB image by combining object detection, monocular depth estimation, 3D model retrieval, and scene assembly. The resulting environment can be explored and interacted with in Panda3D, creating a low-fidelity simulation from an ordinary photograph.

1. Detect objects in the image with a ResNet-based detector.
2. Estimate depth for each object with MiDaS to place it in 3D.
3. Query Sketchfab for candidate 3D models matching each detected class.
4. Human-in-the-loop selection — cycle through matches and choose the best fit.
5. Assemble the scene in Panda3D as a low-fidelity interactive simulation.

Detected object in the source photo → matched 3D model from Sketchfab.

Undergraduate laboratory course providing practical experience with mobile robotics through the development of an autonomous quadrotor software stack. Coursework included dynamics simulation, mapping, state estimation, motion planning, and autonomous exploration.

Graduate-level course on mobile robotics covering the theory and implementation of autonomous systems. Topics include aerial robot autonomy, software engineering for robotics, probabilistic algorithms for perception and planning, and the development of a complete autonomous quadrotor exploration system through five hands-on projects

Mentored a team of five students in a four-week pre-college research program, guiding them through the development of a speech emotion recognition system, from preparing the data and training a machine learning model to deploying it as a web application that runs entirely in the browser using TensorFlow.js.

Combined four public speech emotion datasets (RAVDESS, CREMA-D, TESS, and SAVEE) into a unified six-class dataset, converted audio recordings into mel-spectrograms, trained a convolutional neural network (CNN) to classify emotions, and deployed the final model for real-time, inference directly in the browser.