Experience
Relevant Experience
- Built a camera-LiDAR fusion pipeline for 3D Gaussian Splatting reconstruction of the Mcity test facility in NVIDIA Omniverse NuRec, applying diffusion during training and novel-pose rendering to refine underconstrained regions
- Strengthened response stability and added dataset deletion to an MCP-based agentic LLM interface over auto-labeling and fine-tuning workflows for AV perception
Mapless AI, a Fort Robotics Company
Jan 2026 - Present
Student Fellow, via Perot Jain Tech Lab at Mcity Program
- Engineering a vision-based vehicle platooning system using ROS2 to maintain a safe following distance between vehicles
- Developing an AprilTag perception pipeline with vehicle control systems for deployment on a physical Mustang Mach-E
Michigan Mars Rover
Aug 2025 - Present
Perception Team Member 3rd at the 2026 Canadian International Rover Challenge
- Developing a real-time OpenCV pose estimation algorithm to enable autonomous typing via inverse kinematics
- Optimizing multi-tag 6-DoF pose tracking via image preprocessing, maintaining ≤5mm translation and ≤1° rotation error
- Programmed complex robot behavior in Java by implementing a hierarchical control and command system.
- Integrated data from odometry, inertial sensors, and vision to understand the robot and environment state.
- Utilized frameworks such as VueJS and Typescript to debug the user interactions and data processes in Fusion.
- Created quality assurance checklists for company industrial projects, Fusion and Defender.
Tandon Machine Learning Summer Program
Aug 2022
Machine Learning Student
- Built a car recognition machine learning program using transfer learning with 95.1% testing accuracy.
- Experimented with various conventional machine learning transformation techniques to classify images.
- Minimized overfitting to the training data by augmenting the data.
Research
Lab Work
ROAHM Lab at the University of Michigan
June 2024 - Present
Research Assistant, advised by Prof. Ram Vasudevan
- Investigate hierarchical motion planning algorithms that combine risk-aware trajectory generation with nerual reconstructed environments (Gaussian Splats and NeRFs).
- Develop and integrate a method for estimating the forward reachable set (FRS) of a quadrotor using spherical approximations to enable a safety-guaranteed trajectory optimizer.
- Design a high-level planner by applying an existing probabilistic collision model to generate strategic waypoints, improving the efficiency of the low-level FRS-based planner.
- Co-authoring paper on obstacle avoidance for quadrotors (accepted to IROS 2026).
Publications
AirSplan: Risk-Aware Motion Planning for Quadrotors in Cluttered 3D Gaussian Splats
Seth Isaacson, William Hong, Katherine A. Skinner, Ram Vasudevan
IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2026
Collision-free quadrotor navigation that combines normalized 3D Gaussian Splatting scenes with a reachability-based motion planner, using differential flatness to derive tight continuous-time collision constraints in cluttered environments.
Projects
Selected Projects
Mcity DriveStudio
June - July 2026
3D Gaussian Splatting reconstruction of the Mcity test facility, with a diffusion model.
PythonGaussian SplattingDiffusionNVIDIA
- Modified a camera-LiDAR fusion pipeline for 3D Gaussian Splatting reconstruction of the Mcity test facility.
- Integrating NVIDIA's Difix3D+, a diffusion model, during training and novel-pose rendering to enhance underconstrained regions.
UMTRI Parsivel Dashboard
June - Aug 2026
Website that displays and visualizes precipitation sensor data from a database.
Vue 3VuetifyFlaskSQLAlchemySQL Server
- Created website that shows a landing page, a table of the measurements, a hyetograph, and a weather-type chart. The website also supports export filtered data as a CSV.
- The dashboard takes in data from an OTT Parsivel² laser disdrometer located outside the University of Michigan Transportation Research Institute (UMTRI) building.
Differentially-Flat Aerial Manipulator Simulation
Apr - May 2026
Trajectory reconstruction for an underactuated aerial manipulator in MATLAB via differential flatness and Lagrangian reduction.
MATLABDifferential FlatnessLagrangian Mechanics
- Built a kinematic simulator for a quadrotor with a 2-DOF arm, based on the differential-flatness method in arXiv:2111.01302.
- Recovered full base pose from a commanded center-of-mass trajectory with algebraic and integration-based methods.
Lindenmayer Systems
Sept - Nov 2025
Procedural 3D tree generation via L-systems with zonotope branch wrapping.
PythonL-SystemsProcedural Generation
- Developed a procedural tree generation system using Lindenmayer systems (L-systems) to create three-dimensional botanical structures with physically accurate branching patterns.
- Created a custom L-system interpreter with rotation operators that parses generated character strings into geometric branch endpoints.
- Designed a collision geometry representation by wrapping each branch segment in n-gonal prism zonotopes with thickness parameters that decrease proportionally with distance from the trunk.
- Optimized the zonotope set through combining collinear zonotopes to reduce computational overhead for computer vision applications.
RRT Algorithms
May 2025
A path planning algorithm that efficiently navigates around obstacles, with an optimizer to automatically tune its parameters.
PythonSampling-Based Path PlanningBayesian Optimization
- Developed the Rapidly-Exploring Random Tree (RRT) algorithm for efficient path planning in obstacle-rich environments.
- Implemented biased and angular node steering constraints to accelerate path discovery, along with live visualization of tree expansion and final path generation.
- Built a Bayesian Optimization pipeline to find optimal RRT parameters, featuring cumulative session logging and regional analysis to identify robust parameter ranges.
Other Projects
Mcity LiDAR to IMU Calibration
July 2026
Targetless LiDAR to IMU calibration using OA-LICalib, modified for Mcity.
PythonCalibration
Mcity Camera to LiDAR Calibration
July 2026
Target-based MATLAB camera to LiDAR calibration with MATLAB/OpenCV camera intrinsic calibration for Mcity.
MATLABPythonCalibration
- Addressed the steep learning curve of the Linux terminal by designing a desktop app that translates natural language queries into executable commands.
- Developed a full-stack system using Electron.js and a WebSocket backend, collaborating with a team to integrate the Gemini API for remote command execution.
Education
Academic Background
University of Michigan
2025 - 2028 (exp)B.S Computer Science Engineering & Minor in Mathematics
University of Michigan
2025 - 2028 (exp)
Cumulative GPA: 3.98/4.0
Current Relevant Coursework:
- EECS 445: Machine Learning
- EECS 370: Computer Organization
- Math 425: Probability
Completed Relevant Coursework:
- EECS 281: Data Structures and Algorithms
- EECS 280: OOP in C++
- Math 316: Advanced Differential Equations
- Math 217: Linear Algebra (Proof-Based)
- Math 285: Honors Multivariable Calculus
- Math 465: Combinatorics
- Math 185/186: Honors Calculus I & II
Activities:
- ROAHM Lab
- MRover
- Perot Jain Techlab
University of Michigan
Aug 2024 - May 2025Dual Enrollment
University of Michigan
Aug 2024 - May 2025
Relevant Coursework:
- Math 217: Linear Algebra (Proof-Based)
- Math 285: Honors Multivariable Calculus
- Math 185/186: Honors Calculus I & II
Saline High School
2021 - 2025Valedictorian
Saline High School
2021 - 2025
Cumulative GPA: 4.0/4.0
Relevant Coursework:
- AP Computer Science A
- AP Physics C: Mechanics
- AP Physics C: E&M
- AP Calculus BC
- AP Statistics