My research at the Existential Robotics Lab (UCSD) involves reinforcement learning for manipulation of an xArm6 robot. I am currently exploring imitation learning from image-based demonstrations.

Robot Arm Manipulation

When life gives you Carnegie Mellon, play with robots! Debugged Husarion software on ROSbots to enable mapping and localization for executing path planning algorithms, and collected cool demos!

Path Planning with ROSbots

Solved a generalized version of the multi-agent traveling salesman pathfinding problem under agent-target assignment constraints. Developed an optimal approach, as well as a heuristic offering a 60% speedup with <1% cost increase on average

Multi-Agent Path Planning

Scaling multi-agent communication learning on supervised and reinforcement learning tasks, based on a message encoding approach that allows gradient backpropagation through discrete messages.

Multi-Agent Communication Learning

Tuned large language models (BLIP2, Alpaca LLM) to generate creative captions for images, and designed evaluation methods to score generated captions as well as the custom evaluator model used for scoring captions

Automated Social Media Caption Generation

Tuned a convolutional neural net for multi-output regression to predict collider parameters in nuclear power plants. Achieved a 3x reduction in model size and comparable performance to the best Kaggle model for the task

Collider Detection using Vibration Data

Implemented a variational autoencoder (VAE) and generated medical images based on a chest x-ray dataset from Kaggle

Generating Medical Images using Variational Autoencoder

Addressed missing local features in NeRF-based 3D scene reconstruction using convolutional occupancy nets

Locally-Aware Neural RGB-D Surface Reconstruction

For a given differential drive robot (in simulation) and a world model, generated the configuration space and implemented path-finding algorithms using greedy search, Voronoi diagram, probabilistic roadmaps (PRMs) and rapidly exploring random trees (RRTs)

Robot Path Planning Algorithms

Implemented an approach to estimate the dominant planes from given point clouds of scenes cluttered with objects, by applying the RANSAC algorithm and K-means clustering

Estimation of Plane Parameters from Point Clouds

Investigated the performance of interpolation methods in reconstructing the high-temporal-resolution trajectory of
a self-driving car from waypoints gathered at increasingly lower sample rates

Numerical Analysis for Trajectory Reconstruction

• Scaling up with Graph Neural Networks: Scaled multi-agent coordination from 3 to 15+ agents fornavigation under partial observability with randomly displacing goals using graph neural nets for RL training
• Dynamic Networks and Simulations: Initiated study of low-bandwidth comms networks with interruptionsand employed curriculum learning to speed up training. Developed a visualization framework for comms channelsin OpenAI Gym, accelerating progress in subsequent multi-agent reinforcement learning projects in the lab

Multi-Agent RL with Graph Neural Nets

Developed a novel deep learning approach to OHWR using sequence-models to tackle infinite vocabulary, achieving an accuracy of 80.8% comparable to then SOTA 81.6%

Online Handwriting Word Recognition

Microsoft Summer Internship: Developed a ChatBot to assist the on-call individual in the Azure Data Lake Storage team. Presented project poster at Microsoft's internal machine learning and data science conference (MLADS Synapse).

Summer of ChatBot!