Integrated a DDP smoother into the training/preprocessing pipeline of an end-to-end planning model trained on 1.3B human-driving frames. Shipped an augmented-Lagrangian DDP denoiser cutting jerk/lateral-jerk discomfort rates to 0-4% from 8-39%. Re-architected a legacy C++ production planner’s trajectory-optimization core as an augmented-Lagrangian DDP solver, improving solve speed ~6x over an IPOPT baseline. Designed an elastic, L1-penalized constraint-relaxation scheme shipped onboard.

Investigating inference-time safety mechanisms for diffusion policies in long-horizon robot decision-making. Built a PyTorch training and evaluation framework for Diffusion Transformer policies across navigation and manipulation domains.

Engineered a rapidly deployable autonomous ground robot capable of exploring, mapping, and navigating previously unseen terrain. Developed an interpretable autonomy stack, implemented online mapping and real-time replanning, and developed a Python data access layer over ROS2 bag files. Deployed full autonomy stack on embedded onboard compute.

Proposed a sampling-based weight-space projection framework for enforcing safety constraints during policy updates. Formulated policy optimization as a convex projection and derived sufficient stability conditions. Empirically demonstrated safe performance gains in regression and imitation learning. Oral presentation at IFAC World Congress 2026.

Co-authored an off-policy visual-domain-transfer method for end-to-end imitation learning that aligns source and target latent distributions conditional on vehicle state using a discriminator-based conditional-KL estimator. Demonstrated robust transfer and strong sample efficiency in BARC-CARLA driving environments.

Reviewed seven submissions across Automatica, CDC, IFAC World Congress, IEEE CASE, and IEEE Robotics and Automation Letters.

Designed a constraint-aware imitation learning framework incorporating an actor-critic structure. Improved training efficiency by 150% and reduced collision rates by 60% in CARLA simulation. Integrated reachability-based safety filters. Oral presentation at IEEE IROS 2025.

Led a team of four undergraduate students on vision-based off-road vehicle dynamics modeling and path planning.

Developed a vehicle dynamics simulator based on dynamic bicycle model with gymnasium APIs. Implemented a CARLA connector.

Led discussions and office hours. Developed labs and exams on Python, MATLAB, and numerical methods.

Developed a CNN-based end-to-end controller for high-speed autonomous racing using RGB camera input. Trained using imitation learning and deployed on a 1:10 Jetson-powered vehicle. Achieved long rollouts without constraint violation.

Explored a goal-conditioned, data-driven MPC pipeline learning transition costs between feasible states via imitation learning, using a two-layer hierarchical architecture with short-horizon reference generation tracked by MPC.

Co-implemented a ROS 1 perception stack to identify object locations and types around the rover, and connected perception outputs to an MPC planner and low-level controller on ROS 2 to enable evasive maneuvers.