Projects

Research Projects

In my research at xLab (UPenn), I explore learning-based methods to improve solving time, dynamics modeling, and optimization structure for control problems in safety-critical settings.

Learning-to-Optimize for Mixed-Integer Nonlinear Programming (2025 – Present)

MINLPs arise naturally in hybrid systems and safety-constrained control, but their combinatorial structure makes real-time solving intractable — learning can amortize this cost across repeated problem instances.
Developed the L2O-MISQP framework, integrating neural networks with a trust-region MISQP solver to accelerate solving complex MINLP problems.
Designed an end-to-end hybrid training pipeline using differentiable optimization layers.
Under submission.
Keywords: MINLP, learning-to-optimize, differentiable optimization, trust-region methods

Hybrid Learning-to-Optimize for Mixed-Integer MPC (2025)

MPC requires solving a new MIQP at every control step; the key insight is that these problems share structure across time, so a neural network can predict warm-start integer decisions and cut solve time without sacrificing feasibility.
Proposed a hybrid learning-to-optimize framework for parametric MIQPs, targeting acceleration of mixed-integer MPC.
Integrated neural integer prediction, a differentiable QP layer, and supervised/self-supervised losses to improve feasibility and optimality.
Accepted at Learning for Dynamics and Control (L4DC) 2025. [arXiv]
Keywords: differentiable optimization, MIQP, learning-to-optimize, MPC

Gaussian Process Dynamics with MPPI (2025)

Physics-based vehicle models degrade under changing conditions; a GP learned residual can correct mismatch on-the-fly, but only when the uncertainty is well-calibrated.
Learned GP dynamics from F1TENTH vehicle data and integrated uncertainty into Model Predictive Path Integral (MPPI) control.
Conducted a systematic study of GP-residual dynamics under varying levels of model mismatch, identifying conditions where learned corrections improve performance and where they cause instability.
Keywords: Gaussian Process regression, uncertainty-aware control, MPPI, Bayesian learning

Autonomous Systems

MPC Controller for BeamNG (2026 – Present)

Simulation-to-real transfer is hard; BeamNG provides a high-fidelity closed-loop testbed to stress-test planning and control algorithms before hardware deployment.
Developed a vehicle control pipeline for trajectory tracking and high-speed autonomous driving in BeamNG.
Implemented and evaluated LTV-MPC for tracking human-driven waypoints, achieving 80% of human-driver lap pace on the Spa track.
Ongoing: building Actor-Critic RL with DiffMPC and Vanilla MPPI in the same environment.
Code: GitHub - MPC
Keywords: MPC, LTV-MPC, autonomous driving, reinforcement learning, DiffMPC

F1TENTH Autonomous Racing (2025)

Built a complete ROS 2 stack for perception, planning, and control for autonomous racing.

• Planning & Control
  • Implemented Wall-Follow, Gap-Follow, Pure Pursuit, and MPC controllers with dynamic obstacle avoidance and race-line optimization.
• Perception
  • Developed LiDAR-based localization and vision-based vehicle detection.
• Races
  • Completed two time-trial events and one head-to-head race on hardware.
  • Placed 1st in the first time-trial event; first team to achieve fully headless on-hardware deployment and real-time debugging.

Tools: ROS 2, Python, C++
Keywords: autonomous racing, motion planning, perception, MPC, embedded robotics

LLM-Guided Navigation for F1TENTH (2025)

Natural language is the most accessible interface for non-expert operators — bridging it to low-level control requires grounding semantic intent into safe, executable primitives.
Created an LLM-based planner converting natural-language commands to driving primitives.
Integrated symbolic reasoning with classical control for interactive autonomous-navigation tasks.
Validated in simulation and real F1TENTH environments.
Keywords: LLM robotics, task planning, autonomous driving

Classical Control and Learning Foundations

Quadrotor Control Paper Reproduction (2025)

Reproduced nonlinear quadrotor control using cascaded attitude/position controllers.
Analyzed robustness under parameter uncertainty and external disturbances.
Developed grid-based planners and sensor-fusion routines for trajectory generation.
Tools: Python

Image-to-GPS Regression (2024)

Implemented Vision-Transformer regression to estimate geographic coordinates from street images, benchmarking against classical feature-matching baselines.
Keywords: vision transformers, geolocation, regression