Robotics · Insect-Inspired AI · Computer Vision
Insect-inspired AI for tiny robots.
I am a Ph.D. student in Computer Science at the University of Maryland, advised by Prof. Yiannis Aloimonos. A honeybee crosses miles of changing terrain, finds food, and returns home on roughly a million neurons. I build insect-inspired AI on those terms: compact representations that keep only what a robot needs in order to act, so palm-sized robots can navigate on the power, memory, and compute they actually carry.
Research Objective
One compact perception–memory–action loop.
Hover a brain region to open a research pillar. Tap to pin.
coded aperture · polarizationDesign the optics for the task. CodedVO and PolarDepth encode metric depth in a single image, so one small camera replaces a depth sensor.
computervision·metricdepth·visualodometry
ring S¹ · torus T²Shape memory like the task. Heading is a circle, route progress a torus, homing a single vector, small enough to run onboard.
task-shapedmemory·ringattractor·onboardnavigation
viability · active sensingKnow enough to act. Minimum-Information Viability sets the least sensing needed to reach a goal safely, and says when to move to see more.
informationfloor·activesensing·flapping-wingrobotsNews & media
Full-day workshop in Pittsburgh, 1 October 2026.
Talk at the 2026 Telluride Neuromorphic AI Workshop.
CGTN America, 4 July 2026.
To be presented at IROS 2026.
UMD Department of Computer Science, 2025–2026.
Inaugural AIM Research & Learning Symposium, 5 May 2026.
Research output
A single polarization camera recovers cues that RGB and active depth sensors often miss around glass, improving monocular depth and visual odometry in glass-dominant environments.
A coded aperture physically encodes metric depth cues into images to resolve scale ambiguity in monocular visual odometry.
Roles
University of Maryland
Aug 2024–Present · Dean’s Fellowship · Future Faculty Fellow · GPA 4.0/4.0
Zupt LLC
Jul 2023–Jan 2024Underwater object pose estimation, Sim2Real Blender environments, and dynamic LiDAR simulation.
University of Maryland
2021–2023 · GPA 3.7/4.0Perception and Robotics Group, UMD
Aug 2021–May 2023Drone navigation in unknown environments using aleatoric uncertainty in optical flow, published in Science Robotics, plus 3D vision and active perception.
Ford Motor Company
Jun–Aug 2022Pseudo-object removal, semantic/depth inpainting, HD map generation, and localization.
Swaayatt Robots
Feb–Jul 2021Visual/LiDAR odometry, LOAM, localization, and sensor fusion for autonomous driving.
Portfolio
3D reconstruction, feature correspondence, non-linear PnP, triangulation, and camera pose estimation.
Iterative magnitude pruning on 5% of weights with a 1000-sample dataset, retaining a generalizable model for diverse computer vision applications.
Semantic segmentation across LiDAR and camera using SegFormer for multi-sensor scene understanding.
Image segmentation using superpixels generated with SLIC and k-means, classified with high accuracy using VGG16.
Panoramic image stitching using both classical homography-based methods and deep learning approaches.
Collaborate