Vansh Thakur

Developing a rapid trajectory optimization solver for online constrained motion planning in high-DoF contact-rich systems. Built from scratch, the solver achieves faster solve times with improved numerical conditioning and reduced sensitivity to hyperparameters through online parameter updates. On a double pendulum benchmark, it reduced motor torque usage from 60Nm to 22Nm while satisfying all constraints. Scaled the implementation from MATLAB to C++ for deployment on a Kinova Gen3 7-DoF manipulator, with plans to extend to the Unitree Go quadruped. Ongoing work focuses on improving dynamic feasibility and execution speed.

• Designed dual-layer MPC system for autonomous vehicle simulations in CARLA, replacing unstable PID loops with constrained optimization.
• System encoded human driving data; achieved <64ms rollout and tracking error < 0.3.
• Integrated Gurobi-based MILP backend and linearized QP solver for real-time control in urban environments.Final Presentation

Focused on modeling and control of dynamic systems, emphasizing mathematical modeling and stable control implementation using data-driven methodologies.

• Implemented Model Predictive Control for spring-damper system in C++ and explored data-driven MPC using Gaussian Processes for system identification.
• Developed comprehensive non-linear quadcopter dynamics via Newton-Euler formulation in Python and MATLAB.
• Designed LQR control for trajectory tracking, later augmented with MRAC to handle model uncertainties across multiple parallel systems. View Paper
• Investigated various Deep MPC approaches, implementing tube-based MPC in MATLAB with online deep learning using Radial Basis Function networks.
• Performed comprehensive simulation testing against model uncertainties. View Report | GitHub

Final project for a graduate Linear Feedback Control course. Designed optimal multivariable feedback controller to outperform industry-standard SISO loops for UM's 3x3 RIE plasma chamber process.

• Used LQR + state integrator + Kalman Filter to meet specs: decoupled transient, zero steady-state error, low-frequency uncertainty rejection.
• Reduced voltage spikes by 87% (0.6 → 0.078), halved RF power, and sped up convergence (47s → 11s).
• Process was control redundant — analytically derived limits of SISO strategy and outperformed it using a reduced 2x3 MIMO design.GitHub

From a data-driven control course: learned dynamics of the chaotic Lorenz system using neural approximations of the Koopman operator.

• Trained deep autoencoder in PyTorch to project 3D → 15D latent space; learned linear Koopman dynamics in latent space.
• Loss enforced in 3D space for interpretability; achieved testing MSE ≈ 0.4 across 60k samples of chaotic evolution.

Developed full autonomy stack for a differential-drive mobile robot using ROS2, SLAM, localization, and motion planning.

• Created occupancy grid mapping from scratch using RPLidar data in C++.
• Implemented Monte Carlo Localization with particle filters and hand-coded motion/sensor models.
• Used A* path planning to successfully achieve autonomous maze navigation.
• Performed intrinsic/extrinsic camera calibration using pinhole model in Python.
• Executed forward/inverse kinematics for pick-and-place tasks; modeled Cobot using URDF with RViz joint visualization.

Designed and programmed a complete ADAS (Advanced Driver Assistance System) stack featuring cruise control and lane keeping, using model-based design and embedded C.

• Built Simulink+C implementation of automatic cruise and lane centering on a physical testbed with a haptic steering wheel.
• Performed stability/response analysis using MATLAB, and programmed control logic bare-metal in C.
• Implemented virtual force feedback (springs, walls) via direct control of haptic motors.
• Worked extensively with CAN, ISRs, ADCs, and motor encoders.
• Our team set the lab record for the fastest ISR: 248ns (Fall 2024).