Sharveshwaran Umashankar

Comparison of system trajectories over time for all three state variables. The combined model (SINDy + Neural ODE) closely tracks the true system, especially in regions where the standalone SINDy model diverges.

Enhancing Sparse Identification of Nonlinear Dynamics with Data-Driven Residual Learning

Developed a hybrid framework that combines the Sparse Identification of Nonlinear Dynamics (SINDy) method with a Neural Ordinary Differential Equation (Neural ODE) trained on model residuals. The approach preserves the interpretability of SINDy while improving prediction accuracy in noisy and sparse data regimes.

A compact neural network was trained only on the residual dynamics left by SINDy and integrated with the original analytic model. Experiments on a nonlinear damped oscillator showed an average 85% reduction in trajectory error compared to standalone SINDy, demonstrating a data-efficient and interpretable method for improving learned dynamical systems.

Click here to view the full paper.

Roll, Pitch and Yaw (Top to bottom) Closed loop response with the synthesized robust controller using DK-Iteration and 20 sampled Uncertain plants

Robust Control of a Quadrotor using H∞ and μ-Synthesis

Designed and simulated robust hover control systems for a quadrotor under parametric uncertainty in mass and inertia. A linearized model of the nonlinear quadcopter dynamics was derived around hover conditions and used to design both an H∞ optimal controller and a robust μ-synthesis controller using DK-iteration in MATLAB.

The robust controller achieved stable attitude and position tracking despite uncertainty and external disturbances, demonstrating strong robust stability and performance guarantees. This project bridged theoretical control design with practical implementation, showcasing the importance of robust control in real-world UAV flight systems.

Click here to view the full paper

Lung Tumor Segmentation using U-Net and Res U-Net

Developed and compared multiple convolutional neural network (CNN) architectures for automated lung tumor segmentation on CT scans as part of the Medical Segmentation Decathlon challenge. Implemented U-Net, Res U-Net, and their modified variants in PyTorch, analyzing performance trade-offs in model depth, parameter count, and convergence.

The modified Res U-Net achieved the best balance between accuracy and efficiency, delivering strong segmentation results with reduced model complexity and faster training. The project demonstrated how architectural optimization and residual connections can significantly improve medical image segmentation performance in limited-data scenarios.

Building a 3D-Printed Quadruped and Developing RL-Based Locomotion Policies

Designed and built a 12-DOF quadruped robot from the ground up using custom 3D-printed actuators powered by BLDC motors and magnetic encoders. The robot was controlled via a Raspberry Pi 4 and field-oriented motor controllers (Moteus R4.11) connected over CAN-FD.

Developed a URDF model for simulation in PyBullet and trained reinforcement learning (RL) policies for pose stabilization and gait generation. Explored sim-to-real transfer through a state-corrector model to reduce the gap between simulation and hardware behavior.

Suspension Design and Analysis for Mega ATV Championship

Led the suspension design team for an all-terrain vehicle competing in the Mega ATV National Championship (E-ATV segment). Designed and built the complete suspension system through detailed vibration, structural, and kinematic analyses, ensuring optimal ground contact, stability, and steering performance across off-road conditions.

The final design achieved superior ride quality and maneuverability, contributing to a 1st-place finish at the National Championship.

ASME Human Powered Vehicle Challenge

Led a team of seven in designing a human-powered vehicle optimized for efficient urban transport. Engineered an aerodynamic fairing achieving a drag coefficient of 0.14 at 60 km/h, while maintaining high maneuverability with a 3 m turning radius. The vehicle ranked among the top 15 teams globally in the ASME Human Powered Vehicle Challenge, demonstrating a strong balance between aerodynamic performance and practical design.