Research & Projects
A collection of my work in robotics, perception, and autonomy.
Senior Year Project (Capstone)
Autonomous UAV for Crop Phenotyping
The Research: Developed an autonomous UAV framework for longitudinal crop phenotyping, enabling high-resolution spatiotemporal analysis of wheat varieties across a multi-month field campaign.
The Implementation: Integrated a Livox MID360 LiDAR and Arducam RGB payload with an Odroid N2+ companion computer on a custom 500mm quadrotor. Deployed a ROS-based mission interface for autonomous raster flights, utilizing the FAST-LIVO2 C++ stack for robust LiDAR-inertial state estimation. The pipeline generated field-aligned structural representations and high-fidelity 3D maps, supporting localized temporal canopy-height comparisons evaluated to centimeter-scale accuracy against manual ground truth.
Research & Engineering
Safe Multi-Agent RL for Navigation
The Research: Developed a fully self-contained Python simulation environment for cooperative multi-robot navigation, replacing standard ROS/Gazebo setups. Trained a decentralized Gaussian actor and centralized attention-based critic using Soft Actor-Critic (SAC).
The Implementation: Engineered the continuous-time simulation environment, the simulated LiDAR perception pipeline, and a real-time Model Predictive Control (MPC) safety module using ACADOS. Implemented a three-stage curriculum learning strategy, enabling the agents to achieve an 89.4% success rate while navigating static obstacles under hard safety constraints.
Stochastic Multi-Robot Task Allocation
The Research: Designed an integrated multi-robot warehouse coordination framework combining stochastic utility-based task allocation (SCOBA) with spatiotemporal planning (PM-CBS) over topometric region graphs.
The Implementation: Developed a continuous warehouse simulator in Python to benchmark task allocation algorithms. Implemented an Extended Kalman Filter (EKF) utilizing noisy GPS and ray-cast LiDAR models to reliably track planned paths under state uncertainty. The SCOBA+PM-CBS architecture guaranteed conflict-free execution with exactly zero collisions across all runs, achieving a significantly higher utility-per-distance efficiency (3.19) compared to greedy A* and Q-learning baselines.
Visual-to-Motion Robotic Sketching
The Research: Designed a real-world vision-guided robotic drawing system bridging the sim-to-real gap by combining deterministic continuous trajectory generation with live visual monitoring.
The Implementation: Engineered an image-processing pipeline using OpenCV to extract 2D contours and scale them into 3D Cartesian waypoints. Actuated an Interbotix WidowX 250s arm using ROS (Noetic Ninjemys) and MoveIt for physical execution. Developed a closed-loop monitoring pipeline utilizing an Intel RealSense D455 camera and HSV thresholding to quantitatively evaluate stroke length and execution accuracy on the physical medium.
Incremental Structure from Motion (SfM)
The Research: Implemented a complete pipeline for Incremental Structure from Motion (SfM) to recover 3D geometry from 2D image sequences. Applied multiple-view geometry fundamentals, including SIFT feature extraction and a hybrid geometric verification strategy utilizing Homography and Fundamental matrix constraints.
The Implementation: Engineered an incremental reconstruction loop using Perspective-n-Point (PnP) and implemented a custom Bundle Adjustment step via sparse Jacobian optimization to minimize drift. Successfully generated a dense point cloud of over 88,000 points and integrated the photogrammetry data into a custom Three.js interactive viewer utilizing view-graph navigation and spherical linear interpolation (Slerp) for seamless camera transitions.
Real-Time Age & Gender Prediction
The Research: Formulated a dual-head deep learning framework utilizing ResNet architectures to simultaneously execute age regression and gender classification. Integrated Bayesian techniques via Monte Carlo dropout to quantify predictive uncertainty for real-world reliability.
The Implementation: Optimized and exported the trained models to TensorFlow.js to enable real-time, server-independent inference directly on edge computing devices. Managed the deployment trade-offs of compressing the model precision from 32-bit to 16-bit floats to ensure low latency, ultimately achieving a gender classification F1 score of approximately 94% and an age estimation mean absolute error (MAE) of 5.17 years.
Technical Documentation & Teaching
Lab 9: Perception-Action & Sim2Real
The Concept: Authored curriculum guiding students through building perception-action pipelines using the Interbotix WidowX 250s manipulator. Covered spatial scaling, TF2 transformations, and Cartesian path planning.
Hardware Deployment: Developed a supplementary Sim2Real module addressing physical deployment challenges such as friction and lens distortion. The module instructs on dynamic vision integration utilizing RealSense camera calibration and 6-DOF AprilTag fiducial tracking.
Lab 10: Autonomous Manipulation
The Concept: Designed an advanced laboratory module focused on 3D spatial reasoning and logic-based manipulation for vision-based block sorting and stacking.
Hardware Deployment: The manual guides students through complex trajectory constraints, including dynamic Z-height calculations for tower building, grasp symmetry optimization using modulo operations, and strategies for avoiding kinematic singularities during manipulator execution.