I build learning-based systems for real-world robot manipulation
from visual observations and minimal demonstrations.
I am a Robotics Software Engineering Intern at AgRobotics, working on perception and VLA training data for a dual-arm strawberry harvesting robot. I completed my M.S. in Robotics at the University of Minnesota, advised by Prof. Karthik Desingh at the Robotics Perception and Manipulation Lab, where I published first-author work in IEEE RA-L on category-level last-meter navigation.
My long-term research goal is to develop generalizable robot manipulation systems that perform everyday tasks under severe data constraints. I am particularly interested in how unifying vision, representation, and action can push manipulation beyond highly structured and stationary settings.
Research Directions
Data-Efficient Imitation Learning
I study how robots can generalize from very few demonstrations by reasoning about correspondence between what they were shown and what they currently see. In my RA-L work, a dense score matrix between goal and current observations let a policy trained on a single object instance generalize across the entire category. I am now interested in extending this concept toward in-context imitation, where a policy maps a reference demonstration onto the current scene to decide the next action, serving as the low-level policy beneath a high-level reasoning model.Mobile & Manipulation
For robots to become part of everyday life, knowing how to complete a task is not enough; they also need to move through human spaces and position themselves to carry it out. Most manipulation policies assume a fixed base, while navigation models cannot guarantee the precision that manipulation needs to begin. My RA-L work addresses this gap through last-meter navigation, bringing a robot from the vicinity of an object into a precise, manipulation-ready pose. Next, I am interested in unifying last-meter navigation and manipulation on a humanoid platform as a single mobile manipulation system, while maintaining the performance of the manipulation policy.
What I Build
Real-world data pipelines
Systems that produce training data for imitation learning and VLA policies, including automated demonstration collection on Boston Dynamics Spot using AprilTag-based localization instead of teleoperation, VLM auto-labeling, and curation of noisy teleoperation data by analyzing robot states and actions to identify grasp events and isolate clean demonstrations.Perception systems on robot hardware
Vision pipelines built on foundation models such as DINOv2 and SAM2 for policy learning on Spot, and real-time object detection deployed on NVIDIA Jetson Thor within a ROS 2 stack.Robot software infrastructure
Libraries that make real robots usable for learning research, including SpotStack for perception, data collection, and policy deployment on Spot, and a manipulation toolkit with camera-to-robot calibration and grasp primitives built for a robotics course.
Background and Experience
I completed my undergraduate studies in Mechanical Engineering, which provided a strong foundation in the physical principles underlying robotic systems, including dynamics, control, and state estimation. During this time, I served as the software lead of an Autonomous Underwater Vehicle (AUV) student team, where I was responsible for the design and integration of the software stack spanning perception, estimation, control, and navigation. This role involved extensive work with embedded systems and low-level programming, and gave me hands-on experience deploying autonomous robotic systems in real-world operating conditions rather than purely simulated settings.
In my master’s studies in Robotics, I shifted my focus toward learning-based robotic systems. I initially concentrated on robot perception, particularly vision, through advanced coursework in Robot Vision and Computer Vision, where I implemented projects on visual servoing and monocular visual SLAM. Building on this perception foundation, my master’s research explored learning-based manipulation and robot positioning. Through these works, I became interested in tightly integrating perception, representation, and action to enable robots to perform complex tasks in real-world environments under limited demonstrations and visual-only observations.
Contact Me
I am open to PhD opportunities and research roles in robot learning, manipulation, and real-world robotic systems.
Please feel free to reach out if you are interested in my work or would like to discuss potential research directions. I am always happy to chat about research ideas and collaborations that push the boundaries of robotics!