Chung-Pang Wang

I'm currently a second year ECE master student at University of California, San Diego (UCSD), working with Prof. Michael Yip in Advanced Robotics and Controls Lab (ARCLab).

Prior to coming to UCSD, I completed my Bachelor’s degree in Mechanical and Electro-Mechanical Engineering (MEM) at National Sun Yat-sen University (NSYSU), advised by Prof. Chen-Wen Yen and Prof. Keng-Hao Liu.

I am interested in developing high-performance control methods for robot autonomy, emphasizing physically and mathematically consistent formulations, data-efficient learning, and optimal control schemes to enhance robots' capabilities.

Email  /  CV  /  Scholar  /  Github

Publications

† indicates equal contribution.

We propose a framework for autonomous surgical robotic assistance that integrates a differentiable physics model with perceptual feedback. Our approach achieves two key objectives: 1 ) maximizing tissue exposure and applying tension at a specified dissection site using visual-servoing control, and 2 ) selecting optimal control positions for the dissection target based on deformable Jacobian analysis. We validate our method through repeated real-robot experiments on a tissue phantom and demonstrate its capabilities in dissection experiments using shared autonomy on real animal tissue.

We developed a novel facial asymmetry measurements framework to detect Alzheimer’s dementia (AD), providing a non-invasive and cost-effective method for early AD diagnosis.

Projects

Tracking camera orientation with IMU data. Formulated a optimization problem with queternions and use gradient descent algorithm (GD) to find optimal orientation sequences. Subsequently, I utilize the optimal orientation to construct panoramic images by stitching images over time.

Implemented a lidar-based SLAM algorithm on a differential-drive robot by integrating multiple sensor inputs—wheel encoder and IMU odometry, 2-D LiDAR scans, and RGBD imagery. The robot’s position and orientation are estimated using odometry and LiDAR measurements, enabling the construction of a 2-D occupancy grid map that captures both free and occupied areas of the environment. Additionally, color information from the RGBD sensor is used to enrich the map, assigning realistic floor colors to the 2-D representation.

Visual-Inertial SLAM using an Extended Kalman Filter (EKF). SLAM fuses data from an inertial measurement unit (IMU) and a stereo camera pair to accurately track the vehicle’s trajectory while simultaneously reconstructing a sparse 3D map of the environment. By combining the IMU’s motion predictions with stereo camera observations of visual landmarks, the EKF continuously refines estimates of both the vehicle’s pose and the map’s structure.

Formulate Door & Key problem as a Deterministic Shortest Path problem (DSP) and solve it with Dynamic Programming algorithm by minimizing the cost of reaching the goal.

Implemented search-based (A*) and sampling-based motion planning (RRT, RRT*, Bi-directional RRT) algorithms in 3-D Euclidean space. Test both search-based and sampling-based motion planning algorithms in 7 different environments with axis-aligned bounding boxes (AABBs) obstacles.

Developed a safe trajectory tracking algorithm for a ground differential-drive robot by formulating a discounted infinite-horizon stochastic optimal control problem. Solved the problem using both certainty equivalent control (CEC) and generalized policy iteration (GPI).

Employed Graph Neural Networks (GNNs) to capture the positional residuals between the XPBD simulation and the observed real soft-body. The learned residual model predicts the positional residuals and compensate the sim-to-real gap, enabling more accurate state estimations of future deformations. Green points are the real point cloud of a thin-shell. The grey mesh is the XPBD simulation compensate with learned residual.

Trained an apart-based SE(3)-equivariant neural network, to estimate Panda Arm’s pose from point cloud inputs. Aimed to enable data-efficient transfer of motion tasks between robotic arms.

Enhanced keypoint detection by integrating attention mechanisms into a VGG19-based encoder-decoder network. Our method improves camera-to-robot transformation estimation from single images.

Simulated a swinging rope using Extended Position-Based Dynamics (XPBD) with distance constraints, ensuring stable, realistic motion through constraint-compliant updates.

A Visual hazard detection system for cyclist safety. Intergrated Yolov4-tiny and DeepSort to track vehicles behind and predict potential hazards for cyclists.

Forked from Jon Barron's public academic website