My primary focus at Cobot is to design and integrate the second generation of the Proxie robot (not released yet). I work across multidisciplinary functions including hardware/software integration, sensor and motor bring-up and calibration, software and test development, system debugging and analysis, new product development, and EtherCAT.

Role

Integrate sensors, actuators, and software to enable the physical intelligence of the next Proxie platform.

We designed the world first construction system that is capable of automatically designing, manufacturing and assembling the train tracks, derived from the commercial Brio train tracks.

Deng S, Cao S, Elwin M, Rubenstein M. Automated Construction: An Integrated Pipeline for Environment-Adaptive Design, Fabrication, and Assembly with Constraints (pending publication)

The goal is that the construction robotic system can autonomously sense all the existing tracks and obstacles, generate custom train track pieces that satisfy constraints such as curvature constraints, manufacture the desired track piece with a 3D printer, and finally place the fabricated train track piece in the corresponding location in the environment. The process is repeated until all the track pieces are connected to form a loop track. To further verify the successful construction of the train track loop, the robot places the original Brio train on the track and the train should drive on the loop track without derailment.

The following components make the system possible to automate the entire train track construction process fully:

• Scanning: Computer vision system to distinguish and locate existing train tracks in the environment and obstacles.
• Design: The planner with constraints (such as minimal curvature constraint) solves and finds the sub-optimal path.
• CAD Generator: Convert the sub-optimal path into CAD files, slice the CAD file into printable files, upload it to the printer, and initiate the print automatically.
• Fault Detection: In case of failure (such as print failure), the system attempts to recover and resume the construction.
• Assembly: The robotic arm grasps the printed part and places it in the corresponding locations.

The goal is to design a system integration framework that enables the robotic platform to collect and analyze soil samples efficiently in unknown environments. The main challenge is to integrate all the subsystems together into a complete robotic package that satisfies the mission requirements.

HEBI Bot is a high-end environment monitoring and mitigation platform that we developed at Robomechanics Lab in collaboration with HEBI Robotics and a top US energy company.

The system is equipped with:

• HEBI motors for actuation and mobility
• Velodyne LiDAR for mapping and navigation
• Cameras for teleoperation
• PXRF sensor for contamination detection
• GPS module for localization

S. Deng, S. Wang, H. Wang, I. Krause, N. Sihota, T. Hoelen, G. V. Lowry, and A. M. Johnson, “Autonomous efficient soil sampling with a ground-based robotic system,” Master’s thesis.

Vivek Thangavelu, Hairong Wang, Shane Deng, Sean Wang, Ian Krause, Thomas Hoelen, Gregory V. Lowry, Aaron M. Johnson, “Efficient Autonomous Soil Characterization with a Ground-Based Robotic System,” in preparation for submission to Journal of Field Robotics.

Joe Norby, Sean Wang, Hairong Wang, Shane Deng, Nick Jones, Akshit Mishra, Catherine Pavlov, Hannah He, Sathya Subramanian, Vivek Thangavelu, Natasha Sihota, Thomas Hoelen, Aaron M. Johnson, Gregory V. Lowry, “Path to Autonomous Soil Sampling and Analysis by Ground-based Robots,” Journal of Environmental Management.

The algorithm is developed to be a more efficient way of taking samples in an unknown environment. The algorithm is implemented in Python with the scikit-learn machine learning library. The performance of the adaptive sampling algorithm is compared with the Boustrophedon algorithm and the performance graphs are generated with the MATLAB Python library. To ensure the fairness of comparison, both adaptive and Boustrophedon algorithms use the same underlying Gaussian Process settings. To compare the accuracy of the reconstruction, the Wasserstein distance is used. The Wasserstein distance is also known as Earth Mover’s Distance (EMD).

Torsobot was the discrete wheeled platform I developed mainly to study different walking gaits at the Biomechatronics, Assistive Devices, Gait Engineering, and Rehabilitation Laboratory at the University of Wisconsin–Madison.

The system is equipped with:

• Raspberry Pi 3B for computation
• Arduino Nano for sampling the encoder
• IMU sensor for pose estimation
• DC motor for main actuation
• Servo motor for active kicker mechanism (not pictured)

Here is our lab video that featured the robot I developed: YouTube (also the first slide in the carousel above).

You can read more about the project on the BADGER lab research page.

Gummy Bear Detector

The analog circuit uses fundamental elements such as op-amps, resistors, and shift registers. The detector identifies gummy-bear color with simple spectroscopy: three colored LEDs shine through the sample and a phototransistor measures light passthrough. A comparator stage filters the signal and drives indicator LEDs for up to four colors: red, green, white, and yellow.

Down the Clown

A replica of the classic arcade game built on the STM32F4 with hand-designed circuits—no HAL libraries; peripherals are configured by writing register bits. The GUI is implemented in Processing (Java).

We design simple robots that can move in any direction on a lattice; each unit is driven by a single actuator. The work studies collective behavior in simulation and develops algorithms to control the swarm.

The simulator is a testbed for control strategies, algorithms, and mechanical concepts in both 2D and 3D.

Tools

• MuJoCo
• Box2D
• Unity
• Taichi (GPU-accelerated)

A 3D survival first-person shooter built in Unity with game-controller input. The player survives six waves of increasing difficulty.

Player actions

Shoot, jump, move, and reload.

Enemies

Spawn randomly, chase the player, deal damage, and can jump between platforms.

Built for Build 18 at Carnegie Mellon: an automated GO board inspired by the classic game, combining mechanical design, sensing, and game logic.

Highlights

• Scissor mechanism to clear arbitrary stones
• Gantry to place pieces
• Camera-based piece localization
• Controller enforcing GO rules

Awards

• NSA Most Innovative — Second Place
• Officers’ Choice Award

Build 18 project gallery

7-DoF arm planning

Given start and goal poses with obstacles, the arm is planned with algorithms implemented from scratch:

• Probabilistic Roadmap (PRM)
• RRT, RRT*, RRT-Connect

2D pursuit planning

The pursuer minimizes time-to-capture against a known agent trajectory on a known map using:

• 3D A* (x, y, t)
• 3D backward A* with 2D forward A* (improved heuristic)

Big Brain Pac-Man

An OpenGL Pac-Man base extended with our planners. LRTA* runs on Pac-Man in a partially observed world: the green dot is a randomly spawned pellet; the red/orange ring is Pac-Man’s vision. Experiments vary ghost count (1–3), ghost policies (A* vs default), and vision radius.

Thermo Fisher Scientific

Led the intern team that designed and integrated a multi-color status light bar into the Nicolet Summit FTIR Pro spectrometer (launched commercially in 2019). Recognized for failure analysis and software testing.

Full video of the feature on YouTube

Enerpac Tool Group

New product development co-op on the E-Pulse line during launch: verification testing, failure analysis, and future-capability work. Identified and resolved a critical controller-board issue before release.

Implementations of learning, filtering, and estimation for coursework and projects in simulation and on real-world datasets.

Areas

• Reinforcement learning (online/offline): EXP3, UCB, Q-learning, value iteration, policy iteration, and related methods
• Estimation: Kalman filter, EKF, particle filter, Bayes filter
• Machine learning: naive Bayes, EM, k-NN, neural networks, boosting
• Deep learning: GANs, CNNs, MLPs, RNNs
• Active learning: ergodic control, infotaxis

Control laws implemented and evaluated in simulation environments such as Webots.

Controllers

• Pole placement
• PID
• Feedforward
• Model predictive control (MPC)
• Model reference adaptive control (MRAC)
• Linear quadratic regulator (LQR)

During my time at Carnegie Mellon University, I became deeply involved in entrepreneurship—ranging from startup classes to participating in venture business competitions.

One of our class projects spun off into something bigger: Sensify Inc. We completed the Ascender Incubator Program in Pittsburgh and are now initiating our pilot study with the Pittsburgh International Airport.

Venturecat 2025: Finalist & First Place in Sustainability and Energy. Read the results.

Read more about Sensify: sensifyrecycling.com.

Throughout my undergraduate and graduate studies, I was heavily engaged in mentorship programs—coaching other students on succeeding in academics and life.

Leadership & teaching

• Northwestern PhD Recruitment Chair
• ME233 Electronics Design — Teaching Assistant
• Center for Robotics and Biosystems Coordinator
• 16778 Mechatronics Design — Teaching Assistant
• 24677 Modern Control — Teaching Assistant
• ME439 Intro to Robotics — Teaching Assistant
• CP125 Engineering Seminar — Teaching Assistant
• UW Madison Undergrad Learning Center — Tutor (CAD / Circuit)

During my junior year, I applied to be a house fellow (Resident Assistant, RA). After interviews, I was honored to be selected as the House Fellow for the Sullivan Residence Hall.

Throughout the year, I managed around 70 students, planned events, performed duty rounds, and promoted inclusivity in the community.

Student Today Leader Forever was a national organization with university chapters across the nation. The main goal was to support communities that needed help. One of the flagship programs was Pay It Forward Trip—a 9-day spring break volunteering trip across multiple states.

During my two-and-a-half-year involvement, I served as both a participant and leader in the UW-Madison STLF Chapter.

Read more about STLF.

Outside of school, I engaged with the community through volunteering projects.

I volunteered with organizations such as Urban League (STEM tutor) and Volunteering Solutions (international volunteer in Thailand).

More recently, I joined the Pre-Scientist Pen Pal program to help more people learn about STEM.

Life’s too short to play it safe—I dive headfirst into new adventures, push boundaries, and soak up every experience I can.

I enjoy

• Hiking
• Tennis
• Rollerblading
• Scuba-diving
• Swimming
• Running
• Muay Thai

I’ve also started social media channels such as YouTube, Bilibili, and Rednote. One of the highest viewed videos reached over 110K views.

Throughout college, I was heavily involved in engineering organizations and various competitions such as Wisconsin Robotics, Hyperloop Challenge, and Baja SAE, working at the intersection of hardware and software.