Yanheng ZhuMy team
NYU Shanghai
Robotics Club · RoboMaster
I lead our RoboMaster algorithm team, working on sentry perception, navigation, embedded control, and decision-making.
02 / Projects
Projects
Robot systems and engineering projects spanning perception, navigation, real-time control, and decision-making.
Selected work
Five case studiesRoboMaster
Team & collaborationsCollaboration
NYU Tandon
VIP ARC Robotics
We are helping ARC build its robots. Their team is also using my open-source sentry software.
Joint team
The Great Lakes
Preparing for DJI RMUC 2027
We collaborate with university robotics teams through The Great Lakes alliance, preparing together for next season.
Universities we're working with
- NYU Shanghai
- University of Illinois Urbana-Champaign
- University of British Columbia
- McMaster University
- University of Nottingham
- National University of Singapore
01RoboMaster · ROS2
RoboMaster Sentry Perception & Navigation
I built the sentry's perception and navigation system, adapting open-source auto-aim to our hardware and competition setup. I connected detection and tracking with our localization, navigation, and deployment software built around Mid-360, FAST-LIO, and Nav2.
02RoboMaster · STM32
RoboMaster Embedded Control & Communications
I developed STM32 and FreeRTOS firmware for the sentry's chassis, gimbal, and shooter. Our largely in-house communications and safety systems connect the onboard computer to real-time motor control, with watchdogs and failsafes.
03RoboMaster · Decision
RoboMaster Radar Fusion & Behavior-Tree Decision
I developed a radar pipeline that aligns Mid-360 point clouds with camera detections, then fuses and tracks targets with a Kalman filter. In the sentry stack, BehaviorTree.CPP turns match state and perception into navigation goals and robot actions.
04Unitree Go2
Monocular Navigation on Unitree Go2
I adapted TinyNav to Unitree Go2 and replaced stereo depth with streaming monocular depth from LingBot. Relocalization, planning, control, and safety stops now run as one deployable ROS2 loop.
05MCM · Problem C
MCM 2026: Audience-Voting Models
For COMAP's 2026 MCM, our team used inverse optimization and LightGBM to study hidden fan votes, elimination rules, and voting-system design. The work received a Meritorious Winner award.
More work
Robotics · ML · Software
· Scene reconstruction
Odin1 Digital Twin
This project builds a digital twin of our lab from measured RGB, LiDAR and motion data. The pipeline aligns point clouds, maps, meshes, and semantic layers, and adapts DirectFisheye-GS for native fisheye Gaussian Splatting.
Demo: full-scene lab reconstruction.
· Object-goal navigation
Odin1 ObjectNav
A single-robot navigation stack derived from my TopoFocus real-world system. It combines VLM-guided exploration and spatial memory with A* routing, TinyNav local planning and guarded chassis control to search for target objects.
The demo shows field runs from the original system. The standalone release still awaits full robot validation.
3DGS · Real-to-sim
Mono2Sim-GS
A calibration-light monocular reconstruction prototype combining cuVSLAM tracking, LingBot geometry, GenWildSplat, and Gaussian Splatting for scene reconstruction and inspection.
Audio ML · PyTorch
Urban Sound Classification
An audio classification ensemble using spectrograms, pseudo-labels, and teacher-student training. It placed 4th in the course's Kaggle competition.
VisualKiwi
A prompt-to-Manim visual learning tool with generated narration, FastAPI streaming, and a browser interface.