I am a .
I'm a Robotics Engineer with a particular focus on the applications of control theory and learning algorithms for building smart & safe autonomous robotic systems. I'm currently working as a Project Assistant at the Wearable Robotics & Control Laboratory, Department of Electrical Engineering, IIT Jodhpur, under the supervision of Dr. Saurav Kumar, where my research revolves around the design & development of efficient controllers for prosthetic powered legs and lower-limb exoskeletons.
I hold a Master of Technology degree in Robotics Engineering from the University of Petroleum & Energy Studies (UPES), Dehradun, India. My M.Tech thesis work develops a Stochastic Nonlinear Model Predictive Control (SNMPC) framework that combines with the existing Regulated Pure Pursuit controller (Macenski et al.) to let mobile robots adapt their path-tracking behavior to their own localization uncertainty and tracking error relative to the reference path, in real time.
My research interests involve integrating techniques from control theory, sensor fusion, computer vision, and machine learning into developing safe & intelligent robotic systems. The kind of robots that particularly fascinates me are Industrial Manipulator Robotic Arms, Field/Service Robots, Drones, Powered Transfemoral Prostheses & Lower Body Exoskeletons, Medical/Surgical Robots, and Unmanned Underwater Vehicles.
8.9
M.Tech CGPA
2
Major Projects
3
Research & Industry Roles
What I have studied
My educational journey that has led me to my current research in robotics.
Faculty Advisor
CGPA
8.9 / 10
Dissertation
"Uncertainty-Aware Trajectory Optimization and Robust Control for Grid-Based Mobile Robots"
My thesis develops a Stochastic Nonlinear Model Predictive Control (SNMPC) framework that adaptively tunes the lookahead distance of a Regulated Pure Pursuit controller, jointly accounting for both the path-tracking error and localization uncertainty. See Projects below for more details.
Dibrugarh University Institute of Engineering & Technology, Dibrugarh, India
CGPA
7.58 / 10
My mechanical engineering foundation — underpins how I approach robot modeling and control today, from deriving state-space equations to designing hardware-aware controllers.
Where I've worked
Research and industry roles spanning wearable robotics, warehouse automation, and software engineering.
Wearable Robotics & Control Laboratory, IIT Jodhpur, India | Supervisor: Dr. Saurav Kumar
PRD Robotech Solutions Pvt. Ltd., Noida, India | Supervisor: Mr. Prashant Kumar Dwivedi
Infosys Limited, Bhubaneshwar, India
My Toolbox
Languages, softwares, and hardware I have used for my robotics projects until now.
C++
Python
MATLAB
ROS 2
Nav2
Gazebo
Simulink
Simscape Multibody
SolidWorks
Linux
Git
GitHub
VS Code
Arduino Uno
Raspberry Pi 5
CubeMars AK80-9 Actuator
Waveshare RS485 CAN HAT
AMS OSRAM AS5048B Magnetic Encoder
LaTeX
Markdown
A Stochastic Nonlinear Model Predictive Control (SNMPC) framework built on top of the Regulated Pure Pursuit (RPP) controller — replacing its heuristic, velocity-based lookahead selection with an optimizer that jointly reasons about path-tracking error and localization uncertainty.
The robot operates in a grid-type indoor map with Aruco Marker-coded nodes for ground-truth position correction, tracking a rectangular reference circuit.
Mean trajectory overlays at v = 0.5 m/s — both controllers track closely at low speed, where uncertainty accumulates more slowly.
Cross-track error at 0.5 m/s, shown with ±1σ bands across 5 runs per controller.
Mean trajectory overlays for RPP (blue) and SNMPC-RPP (red) against the reference path at v = 1 m/s.
Cross-track error along the path at 1 m/s, shown with ±1σ bands across 5 runs per controller.
Mean trajectory overlays for RPP (blue) and SNMPC-RPP (red) against the reference path at v = 1.5 m/s — the speed at which uncertainty-aware control matters most.
Cross-track error along the path at 1.5 m/s: SNMPC-RPP (red) cuts peak error roughly in half versus baseline RPP (blue) at each turn.
40%+ Reduction in MAE
SNMPC-RPP vs. Baseline RPP at 1.5 m/s
50%+ Reduction in RMSE
SNMPC-RPP vs. Baseline RPP at 1.5 m/s
Validated Across :
3 operating speeds - 5 runs/trials each - ROS 2 + Gazebo
The Regulated Pure Pursuit (RPP) controller — a widely used trajectory-tracking algorithm in ROS 2's Nav2 stack — selects its lookahead distance using a simple velocity-based heuristic. This assumes the robot's pose is known perfectly, which breaks down as localization uncertainty (from noisy odometry, IMU drift, and sparse position corrections) grows over time.
I designed an SNMPC-RPP controller that replaces only the lookahead-selection module of RPP, leaving the rest of its architecture (velocity regulation, collision checking) untouched. At each 5 Hz control cycle, it predicts lateral and heading tracking error forward in time (3 time steps in future), propagates the robot's pose-error covariance through that prediction, and solves a constrained optimization (cost function + chance constraints + box constraints on lookahead) using the derivative-free COBYLA method to choose the optimal lookahead distance.
Validated in ROS 2 / Gazebo across three operating speeds (0.5, 1.0, and 1.5 m/s) on a rectangular reference circuit, the controller showed its biggest advantage exactly where it matters most: at higher speed, where uncertainty accumulates fastest, SNMPC-RPP reduced cross-track MAE by over 40% and RMSE by over 50% relative to baseline RPP, while keeping the rest of the robot's motion just as smooth.
Derived a linearized state-space model for an inverted-pendulum-style two-wheeled self-balancing robot, first controlled with PID then LQR, and then optimally tuned the Q & R gains with a multi-objective genetic algorithm to perform position tracking simultaneously while maintaining tilt stability.
Simscape Multibody simulation of the robot balancing under an LQR controller with GA-optimized Q/R weighting matrices.
Simscape Multibody simulation of the robot balancing under an LQR controller with manually tuned Q and R matrices.
Position and chassis tilt angle response of the GA-optimized LQR controller, showing faster convergence and improved tilt stability.
Position and chassis tilt angle response of the manually tuned LQR controller, for comparison against the GA-optimized results.
The same robot balancing under a manually tuned PID controller, for comparison against the LQR/GA approach.
This project implements the classic inverted-pendulum balancing problem: keeping a two-wheeled robot upright while also controlling its position. I derived a linearized state-space model of the robot's position and chassis-tilt dynamics, then built and simulated it in MATLAB/Simulink with Simscape Multibody for realistic rigid-body physics.
I implemented both a PID controller and a Linear-Quadratic Regulator (LQR) for the balancing task, then went a step further: rather than hand-tuning the LQR's Q and R weighting matrices, I implemented a Multi-Objective Genetic Algorithm (MOGA) that searches for Pareto-optimal weightings by simultaneously minimizing the Integral Time Absolute Error (ITAE) for position and for chassis-tilt angle — two objectives that typically trade off against each other.
Comparing the GA-optimized LQR against manually tuned LQR and PID controllers in simulation, the GA-optimized controller reached faster position convergence while maintaining better chassis-tilt stability than either manual baseline — demonstrating that automated multi-objective tuning can outperform hand-tuning even on a well-understood classical control problem.
Dr. Rahul Singhal
Get in touch
If you're just as in love with robots — and the art of building these beautiful beasts — as I am, I'd love to hear from you. Whether you want to collaborate on something, or just nerd out about science, movies, robotics, or space, let's connect!
preetamk97@gmail.com