Research & Projects
My research interest broadly includes multi-agent systems, intelligent autonomous systems, distributed optimization, and aerial robotics. Few of my works are highlighted below.
01
Distributed Optimal Control for Circumnavigation Formation
In this work,we designed a two-stage control protocol to addresses the circumnavigation of a moving target by nonholonomic agents aiming to achieve a desired multi-circular formation.
First, recognizing that some agents may not have access to formation parameters, we employ distributed estimators to enable these agents to estimate the necessary parameters. Second, we introduce a decentralized optimal control algorithm based on a linear quadratic regulator (LQR). This algorithm is designed to minimize a weighted sum of squared deviations from the desired states and control inputs, ensuring that the agents maintain the multi-circular formation.
The proposed method not only ensures the desired formation but also optimizes the control performance by minimizing deviations, making it a significant contribution to the field of cooperative circumnavigation.
02
Optimal Trajectory Generation Using Successive Convexification
Determining the most energy efficient flight path for a UAV in a obstacle filled environment is a complex challenge. In this work, we use successive convexification to address the nonconvex minimum time trajectory generation problem for a 6-degree-of-freedom (6-DoF) quadrotor model. Unlike existing methods that only consider translational dynamics and treat the quadrotor as a point mass, our approach includes both translational and rotational dynamics. By converting the original nonconvex problem into a second-order cone programming (SOCP) sub-problem, we solve it iteratively to achieve an optimal solution.
03
CNet: MultiUAV Communication Architecture
CNet is a cross-protocol communication framework designed for low-bandwidth, multi-UAV applications like search and rescue or ISR. It integrates various protocols to ensure high reliability and Quality of Service (QoS) while reducing network overhead during connection and reconnection, enhancing efficiency and performance in multi-UAV systems.
04
EKF Based Predictive Algorithm for Multi-UAV Formation Control
In this work, we present an EKF based predictive algorithm for leader-follower formation control in UAV swarms. Our method includes detailed insights into both software and hardware frameworks, making hardware deployment straightforward. This algorithm requires low bandwidth for inter-agent communication while enabling fully distributed formation navigation. Its effectiveness is demonstrated through ROS and Gazebo simulations, as well as hardware implementations. This customizable architecture provides a robust platform for aerial robotics .
05
CREST: Constrained Robust Ensemble-based Stochastic Trajectory
Ocean current uncertainty often forces Unmanned Surface Vehicles (USVs) to sacrifice energy efficiency for computational tractability. To solve this, we developed CREST , a framework that handles large-scale ensemble forecasts with a per-iteration complexity independent of the number of ensembles.
Validated across diverse ocean environments, CREST consistently outperforms established methods like STOMP and CoSTA. It provides guaranteed convergence to stationary points and achieves up to a 26× speedup over traditional CVX solvers. Even in extreme current conditions, CREST maintains robust performance, making it ideal for real-time trajectory optimization.​​
06
ML-based Trajectory Optimization for Robot Manipulators
In this work, we developed an ML-based trajectory optimization framework for a UR5 robot using B-spline parameterization and Fully Connected Neural Networks (FCNN). By training the system on data from both simulation and real-robot RTDE interfaces, the model predicts optimal control points for pose-to-pose motion while adapting to variable payloads (1-5kg). The end-to-end pipeline, which utilizes ServoJ commands for execution, demonstrates significant improvements in cycle times compared to conventional planning methods across the workspace.
07
Faster Algorithms Robust Regression
We address the challenge of estimating variables from features corrupted by complex acquisition noise. By introducing functional constraints that bound the prediction loss for each perturbed sample, our framework enables robust estimation without requiring an explicit parametric noise model. The proposed algorithm demonstrates superior performance on large-scale datasets, such as the Bike Sharing dataset with over 240,000 constraints. It achieves a significant 73× speedup over traditional solvers like CVX while maintaining near-optimal accuracy.
08
Vision Based Servo Control of Multi-link Robot Arm
To address performance issues in rigid robot manipulators due to being non-collocated systems, we propose using a vision sensor for direct tip position measurement instead of traditional mechanical sensors. Our work focuses on developing a vision-based method to control the tip position of a planar single-link rigid robot manipulator (SLRM). By integrating feedback from both an encoder and a visual sensor, we enhance tip position control performance. Simulation and experimental results demonstrate that our method outperforms traditional mechanical sensor approaches.
09
AI-based Solutions for Industrial Inspection
To improve precision in industrial inspection, I worked to get a comparative framework evaluating CNN, GAN, and Transformer-based super-resolution architectures to enhance the clarity of low-resolution captures. This research focuses on optimizing the trade-off between perceptual sharpness and computational efficiency, demonstrating that Transformer-based models are particularly effective at capturing the fine structural details critical for automated flaw detection.