Publications

Robotics & Machine Learning:

Latent Activation Editing (LAE) is a framwork for inference-time behavior steering / refinement of pre-trained robot policies without retraining. Through a Latent Collision World Model, the detected unsafe activations are replaced on the fly, delivering 90% fewer collisions (safer behavior) in multi-robot quadrotor navigation while maintaining task success. The work establishes LAE as a lightweight framework for behavior enhancement, deployable even on resource-constrained robotic systems.

WARPD extends diffusion-based policy generation by combining latent diffusion, hypernetworks, and a learned world model. Instead of generating trajectories, WARPD produces policy weights, while the world model provides dynamics-aware corrections during training. This yields reactive closed-loop policies that remain robust under perturbations, scale to longer action horizons, and achieve lower inference cost than standard diffusion policies.

Latent Weight Diffusion (LWD) introduced the shift from trajectory generation to policy generation, where a diffusion model learns a latent trajectory space and a hypernetwork decoder transforms it into policy weights, yielding reactive closed-loop policies. This approach delivers long-horizon robustness and achieves ~45x lower inference compute while maintaining SOTA performance across manipulation and locomotion tasks.

RNBF brings NeRF(Neural Radiance Fields)-inspired 3D scene reconstruction directly into the robot's control loop by learning continuous, differentiable neural signed distance fields (SDFs) online from noisy RGB-D input. Unlike prior methods that are offline or require pretraining, RNBF updates its neural SDFs in real time (5-15 Hz) and feeds them into standard Control Barrier Function (CBF-QP) controllers, enabling safer navigation in unknown environments using only a depth camera.

This work demonstrates the feasibility of using EEG signals for real-time motor intention decoding to control assistive robots. Leveraging Riemannian geometry features from EEG covariance matrices and a lightweight SVM classifier, the system achieves ~70% online accuracy in robot-in-the-loop experiments for left/right arm intention prediction. These results establish a practical pathway toward non-invasive brain–robot interfaces for assistive robotics.

Designed a ResNet based CNN architecture for processing the dashboard camera image data to a perception map, hence producing an innovative Perception-based robust control barrier function using a two-layer DNN.

Soft computing & Energy (Undergrad Research Papers):

Relevant Projects

Developed a multi-robot & multi-target path planning & localization algorithm using deep q learning combined with Graph Neural Network outperforming the Dec-SB and Random Walker algorithm for efficiently solving the Active information acquisition problem.

Developed a real-time Signed Distance Function (SDF) generator in ROS to provide 3D spatial awareness for the Franka Emika Panda robot. Implemented an online continual-learning pipeline from depth images with self-supervised loss and optimized GPU-based reconstruction for integration with a 100 Hz control loop, enabling collision avoidance, motion planning, and adaptive manipulation.

Developed a computer vision-based AI system with YOLO for player and ball detection and a custom CNN for court mapping, delivering real-time analytics to optimize player strategies and identify performance gaps.

Implemented a diffusion-based method using a Markov chain and L1-trained restoration backbone to robustly restore severely degraded facial images with enhanced realism and fidelity, leveraging pretrained models for simplified training.

Developed an end-to-end machine learning pipeline for classifying bone fractures from X-ray images, leveraging MLOps practices with DVC, and deploying solutions on both Azure and AWS. Developed and deployed a Flask-based web interface enabling real-time image upload and fracture detection.

Developed a functional motion planning stack that avoids both static & dynamic obstacles, track the center line of a lane, & handling stop signs. [Behavioral planning logic, static collision checking, path selection, & velocity profile generation] in CARLA simulation.

Developed a Graph Neural Network to learn a distributed policy that mimics the optimal centralized controller considering a multi-agent system with N agents tasked with controlling a dynamical process, while ensuring collision and spread avoidance.

Developed the Pulmonary Embolism Detection model based on CNN (Efficientnet-b0) with a weighted log loss of 0.08 for reducing human delays and errors in detection and treatment of PE from chest CT pulmonary angiography images.

Explored the nexus between movie popularity, factors like runtime, genre, and economic conditions, achieving strong predictive accuracy with Random Forest and XGBoost models, highlighting the potential for genre-specific trends during economic fluctuations.

Developed a forecasting model for predicting COVID-19 cases for 81 countries using DNN and LGBM with an accuracy of 97.6% - 99.8%. Dispelled rumors regarding the weather’s role in COVID-19 transmission; examined and demonstrated that weather had little to no role in the spread of COVID-19.

This project explores the classical computer vision technique (non-deep learning) of converting 2D images into 3D Reconstruction.

This project implements a two-view stereo algorithm to convert multiple 2D view-points into a 3D reconstruction of the scene.

This project provides one of the most simplified implementation of the famous Neural Radiance Fields paper "NeRF Representing Scenes as Neural Radiance Fields for View Synthesis".

To address the challenge of child care during customer service visits, a customer-experience enhancement system is proposed, as a part of ESE 514 : Embedded System final project. The system integrates the dino game, a face-tracking pan-tilt camera, and an LCD and speaker for interactive music engagement. Please refer to our website for all the details.

Education

Additional Education / Online Specialization

• Deep Learning Specialization by deeplearning.ai
• Self-Driving Cars Specialization by University of Toronto
• Machine Learning Engineering for Production (MLOps) Specialization
• Algorithms Specialization by Stanford University
• IBM Data Science Professional Certificate
• Applied Data Science with Python Specialization by University of Michigan
• Business Analytics Specialization by Wharton School of the University of Pennsylvania
• Python for Everybody Specialization by University of Michigan

• Machine Learning by Stanford University
• Google Cloud Business Professional Accreditation by Google Cloud
• Deep Learning Specialization by deeplearning.ai
• Introduction to Programming with MATLAB by Vanderbilt University

I'm also using Jon's website template.