Education
Bachelor of Science in Computer Science & Mechanical Engineering, Northwestern University
Evanston, IL
GPA: 4.0 / 4.0
Freshman Dean's List — High Honors
IBDP Candidate, Shanghai Pinghe School
Graduated on the Principal's List (top 5% of class)
Experiences
Undergraduate Research Assistant, Xenobot Lab
October 2025 - Present
- • Built a DW1000-based localization system with three-anchor triangulation, achieving cm-level positioning at 30 Hz update rate. Solved concurrent SPI bus contention with IMU through interrupt-driven mutex, providing a low-cost alternative to expensive motion capture systems.
- • Co-developed modular robot hardware and RL control with PhD researchers. Designed novel wheel-leg modules and trained locomotion policies (PPO) across diverse morphologies in MuJoCo. Developing wheel-leg hybrid modules for multi-modal locomotion.
- • Received $4,000 in Northwestern research funding to continue this work through the summer; selected by the Hung-Farinelli family as one of the fund's featured projects.
Robotics Engineer, HaptE
September 2025 - January 2026
- • Built fully local inference pipeline on Nvidia Jetson with TensorFlow, achieving <20ms per-frame latency with zero cloud dependency. Delivered as a paid monthly subscription service to warehouse clients.
- • Trained recognition system (YOLO detection, MobileViT super-resolution, OCR, QR/barcode) achieving 98%+ item accuracy on 2k-image custom dataset. Integrated LiDAR for 3D spatial awareness in cluttered environments.
- • Developed multi-edge device orchestration with AWS IoT Core for distributed pick-and-place. Integrated LLM-based task planner for intelligent error prevention and autonomous operation.
Research Intern (Algorithm), Differential Robotics
May 2025 - August 2025
- • Designed sparse-to-dense reward curriculum for obstacle avoidance, improving training convergence 3× over baseline PPO in cluttered environments in IsaacLab.
- • Deployed real-time end-to-end vision model for high-speed drone maneuvers through near-vertical frames at 4+ m/s. Achieved 95% success rate with MobileNetV3 optimized via TensorRT (<5ms latency). Successfully transferred to physical hardware.
Projects
End-to-End High-Speed Drone Navigation System
Computer Vision, Reinforcement Learning, TensorRT, Sim-to-Real
- • Led the full sim-to-real deployment of the paper's end-to-end sensorimotor policy, enabling a physical drone to traverse 85° tilted frames at 4–6 m/s with 95% success and high repeatability; the resulting demo was exhibited at IROS 2025.
- • Optimized MobileNetV3 + DeepLabV3+ on Jetson Xavier NX via TensorRT, achieving ~3 ms inference and >60 Hz closed-loop control.
- • Deployed models developed in Precise Aggressive Aerial Maneuvers with Sensorimotor Policies (Science Robotics) and Whole-Body Control Through Narrow Gaps from Pixels to Action (ICRA).
Dexterous Hand Object Reorientation
Isaac Lab, PyBullet, Dexterous Manipulation, Sim-to-Real
- • Assembled and calibrated a 16-DoF dexterous hand platform; built a PyBullet digital twin for joint-level control tests, policy debugging, and hardware-safe command validation.
- • Trained Isaac Lab object-reorientation policies with randomized object poses, dynamics, and actuator parameters, then deployed them on the physical hand and iteratively tuned sim-to-real reliability.
ME 433 Haptic Feedback Joystick
Embedded Control, Haptics, STM32, Raspberry Pi Pico 2 W
GitHub →- • Designed and built a two-axis force-feedback joystick as the final class project for Northwestern's ME 433: Advanced Mechatronics.
- • Integrated modified servos, magnetic encoders, force sensing, and custom STM32/Pico 2 W electronics.
- • Implemented a 500 Hz closed-loop controller that renders programmable textures and forces through the joystick.
F1TENTH Autonomous Driving
ROS 2, LiDAR SLAM, A*, MPPI, Ackermann Control
- • Built and deployed a ROS 2 autonomy stack on a physical F1TENTH car, integrating LiDAR SLAM, localization, and RViz-based debugging.
- • Combined A* global planning with MPPI local trajectory tracking and Ackermann drive control for closed-loop navigation.
Zombies VS Plants
JavaScript, HTML/CSS, Python, Computer Vision, AI Agents
- • Built a real-time PvP strategy game inspired by Plants vs. Zombies, adding playable plant/zombie sides, local 1v1 gameplay, animated sprites, sound effects, and AI opponents trained from recorded gameplay logs.
- • Integrated a computer-vision physical board interface that detects real-world game pieces and maps their type/location onto the in-game grid; built in 24 hours at WildHacks 2026.
Beatbox-to-Drum Generation
Python, PyTorch, torchaudio, Flow Matching, MIDI, Human Feedback
Demo Website →- • Developed a beatbox-to-drum generation system that converts vocal percussion recordings into symbolic 9-class drum patterns using a flow-matching-based generative model.
- • Built an audio preprocessing/model training pipeline, MIDI/audio tooling, and a human-feedback review loop for ranking outputs, training reward models, and iteratively improving rhythmic alignment and musical realism.
End-to-End RL Drone Navigation System
2025
- • Built an end-to-end navigation pipeline in IsaacLab using deep reinforcement learning (PPO), mapping high-dimensional sensor inputs (depth images + proprioceptive states) directly to low-level motor actions.
- • Implemented curriculum learning & domain randomization (obstacle density, dynamic disturbances, sensor noise) to improve generalization and enable sim-to-real transfer.
- • Designed hybrid reward shaping: integrated Dijkstra-based path priors, directional distance rewards, ESDF collision penalties, and smoothness constraints to stabilize long-horizon flight.
Low-Cost UWB Localization for Rollbot
2025–2026
- • Built DW1000 localization with concurrent SPI bus sharing (UWB + IMU) via interrupt-driven hardware mutex for stable multi-device communication.
- • Rewrote TWR ranging protocol, increasing update rate to 30 Hz and improving accuracy via Kalman-filtered multi-anchor triangulation.
- • Delivered cm-level autonomous navigation for Rollbot (single-actuator spherical robot), offering a $50 embedded alternative to $10k+ OptiTrack systems.
Wheel-Legged Modular Robot with Evolutionary Design (Ongoing)
- • Designing wheel-leg hybrid modules enabling multi-modal locomotion (walking, rolling, climbing) with evolutionary morphology optimization.
- • Training RL control policies (PPO via SB3) in MuJoCo for computer-generated morphologies; conducting sim-to-real transfer to physical robot.