Zhongming (Reaven) Huang

Convex MPC for Quadruped Locomotion

Experimenting with a convex MPC controller in python and simulate in MoJoCo. Code is adapted from Yinghan Sun's implementation of a Mini-Cheetah cMPC controller.

ECE 5280 Optimal System Analysis and Design, 2024

Multi-robot Construction System with Quadruped Robots

Each agent in this system is capable of retrieving, transporting, positioning and fixing minimally processed foam blocks with off-the-shelf T-pins. Instead of operating in the periphiral of the construction site, each agent considers the structure as a graph of traversable paths. This means the robots climb on the structure to build new blocks. Multi-robot collaboration is experimented in simulation with local behavior rules inspired by the TERMES system, with a focus on building with 2x1 blocks and constructing non-lattice structures.

Submitted to 2025 IEEE International Conference on Robotics and Automation (ICRA), 2024

Fast 4-wheel Autonomous Robot

Assembled a 4-wheeled robot with 2 motors and minimal perception / computation. With a 6-DoF IMU, two ToF sensors and one Artemis Nano board, the robot is able to localize and navigate autonomously in a constrained environment.

ECE 5160 Fast Robots, 2024

Instance Segmentation based on Polygon Regression

Traditional instance segmentation algorithms generates pixel-wise predictions to form closed areas for each instance, which may slow down inference speed and provides unintuitive position representation for each instance. This projects aims at encapsulating the instances' position with its center coordinates and vertices as a polygon. For each instance, the vertices are represented in a polar coordinate frame originated from the center coordinates of the instance. This concept is realized upon a commonly used object detection framework to ensure simplicity and speed.

In Journal of Physics, IOP Science, 2023

Excellent Senior Thesis Award, 2023 (Bachelor's Dissertation)

Accurate Small Liver Cancer Detection

Computer aided cancer diagnosis has great potential to help with prescreening of early-stage cancerous lesions. However, the dataset for training a diagnosis algorithm is limited compared to the scale needed. Small lesions of early stage cancers makes it even harder to collect sufficient data. This project aims at training a classification network in a weakly supervised manner, proposing an algorithm with self-attention to automatically compare an unknown CT scan with those labeled. With limited labeled data, this algoritm extracts context features of a small lesion and its nearby area. With these features, the algorithm is able to give probability of whether an unknown image contains a similar lesion.

In Proceedings of 2022 IEEE/ACIS International Conference on Computer and Information Science (ICIS) , 2022

Robotic Grasping for Partially Occluded Objects

In field tasks, a common goal is to securely grasp an object that is occluded by other undesired targets. Also, in field tasks, it is unlikely to have multiple RGB-D cameras to surround the operation site. These together formulates the task as grasping the occluded target with one fixed RGB-D camera, which may require minimal removal of objects hampering a confident grasp. This work proposes an algorithm that decides the best-next-grasp, which minimizes the number of required intermediate grasps and maximizes the success of target object grasping.

In Proceedings of 2022 IEEE International Conference on Intelligent Autonomous Systems (ICoIAS), 2022

Best Paper Presentation Award at ICoIAS, 2022

Quadrotor Drones for Package Delivery and Fire Inspection

Hosted a provincially funded college student innovation project, with a theme of design and control of intelligent quadrotor drones. Package delivery is a topic demonstrating the stability and accuracy of a drone's sensing, control and actuation. We assembled a quadrotor drone, adapted the PID control loop along with the integration of a OpenMV camera, an optical flow and ToF sensors to enable its reliable operation in indoor environment. The task includes automatically take-off and landing, picking and placing objects in designated locations. The fire inspection drone has an Nvidia Jetson Nano with an IMX 219 camera to detect flame in real-time. A light weight flame detection algorithm was dedicated to be executed on board in real-time with less than 2GB RAM, which extracts brightness features of pixels in the temporal domain as inputs of a binary classifier.

In Proceedings of 2022 IEEE International Conference on Pattern Recognition and Machine Learning (PRML), 2022

Best Paper Presentation Award at PRML, 2022

P.R.China Software Copyright, 2022

Provincial Second Prize of the 7th National College Student Engineering Training Ability Comprehensive Competition, 2021