Santosh M. Rajkumar

Thesis: Haptic Rendering in Touch Surfaces Using Multifrequency Electrostatic Actuation.

• Koopman-based Data-Free Real-Time Model Predictive Control (MPC) for Quadrotors
  • Developed an analytical Koopman-lifted model for quadrotor dynamics that preserves control input dimension.
  • Achieved over 100% improvement in translational motion prediction accuracy with a 26% reduction in lifted state dimension compared to state-of-the-art methods.
  • Introduced KQ-LMPC, the first analytically derived Koopman-based linear MPC framework for quadrotor control.
  • Demonstrated tracking performance comparable to nonlinear MPC with ~ 2–4x reduction in computation time.
  • Experimentally validated on a quadrotor platform (Jetson NX + PX4 + Vicon) Video Demo.
  • Developed open source Python Package for KQ-LPMC: GitHub Repo PyPi Python Package.
• Data-Driven Dynamics Learning from Noisy Partial Measurements for Output Regulation
  • Developed a Koopman bilinear model with a neural decoder (KBM-NL), formulated as a Hidden Markov Model for robustness to noise and partial observability.
  • Designed a customized neural Expectation–Maximization (EM) algorithm to jointly identify the KBM-NL dynamics and latent inference distribution from noisy, actuated trajectories.
  • Achieved output regulation via MPC constructed on the learned KBM-NL surrogate model.
  • Demonstrated superior prediction accuracy and stability on Duffing oscillator benchmarks, with strong generalization to unseen trajectories.
• Output Regulation under Noisy Partial Observations using Adaptive Dynamic Programming (ADP)
  • Eliminated need for belief, observer, or model knowledge, enabling model-free output-feedback control for non-linear systems.
  • Designed a critic-only ADP algorithm using a Lyapunov–based Q-function for optimal control learning.
  • Introduced a persistence-of-excitation mechanism via learned derivative feedback, without an initial stabilizing controller.
  • Enforced closed-loop stability during learning using Lyapunov-constrained temporal-difference (TD) learning.
  • Developed a stable on-policy value iteration scheme with no replay memory or large basis functions.
  • Demonstrated output regulation on a cart-pole system.

• Developed in-house finite element (FE) models for large-area touch displays (1D bar & 2D plate) with spring–damper boundaries to study vibrotactile response and mode shaping.
• Designed a multifrequency excitation strategy using Electrostatic Resonant Actuators (ERAs) to achieve localized haptic rendering with minimal actuator hardware.
• Fabricated working prototypes and validated FE predictions experimentally using vibration analysis with > 90% agreement.
• Proposed an energy-based control strategy to position and steer localized haptic feedback across arbitrary surface locations.
• Optimized actuator placement and boundary compliance to enhance elimination of haptic “dead zones.”

• Maintained automation and electrical systems in critical petroleum installations.
• Applied vibration-based monitoring on rotating machinery.
• Led vision-based robotic inspections of storage facilities.
• Developed web applications (JS, Python, AWS) and managed vendor negotiations and SAP-based acquisitions.
• Trained large industrial workforces in safety procedures.