Google scholar profile: Google scholar profile
Google scholar profile: Google scholar profile
Publication List
Shing-Hei Ho, Zohre Karimi, Bao Thach, Alan Kuntz, & Daniel S. Brown , "Reward Learning from Suboptimal Demonstrations with Applications in Surgical Electrocautery, " International Symposium on Medical Robotics (ISMR), 2024 (Best Paper - Finalist, Best Student Paper - Finalist)
Shing-Hei Ho, Bao Thach & Minghan Zhu, "Generative LiDAR Editing with Controllable Novel Object Layouts, " IEEE International Conference on Robotics and Automation (ICRA), 2025
Shing-Hei Ho & Alan Kuntz, "A Model-free Bayesian Optimization Approach to Surgical Retraction Automation," under review at IEEE Robotics and Automation Letters (RA-L), 2025
Bao Thach, Tanner Watts, Shing-Hei Ho, Tucker Hermans, & Alan Kuntz , "DefGoalNet: Contextual Goal Learning from Demonstrations For Deformable Object Manipulation," IEEE International Conference on Robotics and Automation (ICRA), 2024
Bao Thach, Brian Y. Cho, Shing-Hei Ho, Tucker Hermans, & Alan Kuntz , "DeformerNet: Learning Bimanual Manipulation of 3D Deformable Objects, "under review at the International Journal of Robotics Research (IJRR), 2025
My research supervised by Prof. Shandian Zhe on derivative-enabled bayesian optimization with gradient is in progress. I included a brief description about this research in my statement of purpose in the PhD application.
Shing-Hei Ho, Zohre Karimi, Bao Thach, Alan Kuntz, & Daniel S. Brown , "Reward Learning from Suboptimal Demonstrations with Applications in Surgical Electrocautery, " IEEE International Symposium on Medical Robotics (ISMR), 2024 (Best Paper - Finalist, Best Student Paper - Finalist)
My very first first-author publication !!!
We propose to learn a reward function from preferences over suboptimal demonstratioins consisting of latent feature representations of partial-view point cloud observation. A policy can then be learned using reinforcement learning
Our method reduces the need for near-optimal demonstrations and opens the door to surgical policy learning from qualitative human feedback
What if we can generate diverse self-driving data at a low cost? My work is the first framework that enables flexible and controllable generative LiDAR editing that enables generating novel object layouts while maintaining realistic background. Due to the flexibility and controllability, this framework benefits testing and development of self-driving systems.
Shing-Hei Ho & Alan Kuntz, "A Model-free Bayesian Optimization Approach to Surgical Retraction Automation," under review at IEEE Robotics and Automation Letters (RA-L), 2025
[arxiv link coming very soon]
My work takes the first step to automate surgical retraction without relying on a physics model, which enables generalization to tissues of various geometries
Bao Thach, Tanner Watts, Shing-Hei Ho, Tucker Hermans, & Alan Kuntz , "DefGoalNet: Contextual Goal Learning from Demonstrations For Deformable Object Manipulation," IEEE International Conference on Robotics and Automation (ICRA), 2024
[arxiv] [project website]
Our work is the first to tackle the goal shape specification problem in the shape servoing literature.
DefGoalNet learns to generate a goal point cloud given the initial point cloud of the deformable object and a point cloud containing contextual information of the task.
Bao Thach, Brian Y. Cho, Shing-Hei Ho, Tucker Hermans, & Alan Kuntz , "DeformerNet: Learning Bimanual Manipulation of 3D Deformable Objects, " under review at the International Journal of Robotics Research (IJRR), 2025
We develop a novel DeformerNet neural network to learn the robot action that iteratively deforms the object to the desired shape only based on the current and goal partial-view point clouds. A dense predictor is used to intelligently predict where the end-effector(s) should grasp before manipulation.