Yunshuang (Sheyla) Li

I am an incoming Ph.D. student in Computer Science at University of Southern California, where I will work with Prof. Daniel Seita and Prof. Gaurav Sukhatme. I am interested in robot learning for general-purpose robot manipulation.

I will receieve a M.S. degree in Robotics at GRASP Lab, University of Pennsyvania in May 2024. I was a member of Perception, Action, & Learnin (PAL) Research Group, advised by Prof. Dinesh Jayaraman. Previoulsy, I worked with Prof. Qi Dou at Chinese University of Hong Kong (CUHK) for an internship. I received my Honorable B.S. degree in Automation from Chu Kochen Honors College, Zhejiang University in 2022.

Email  /  CV  /  Bio  /  Google Scholar  /  Twitter

Research

My research interests lie in robot learning for general-purpose robot manipulation. During my PhD study, I would like to focus on learning for robot manipulation, exploring topics such as representation learning and policy learning. My ultimate goal is to enable robots to manipulate and learn to manipulate like humans.

(* indicates equal contribution, † indicates equal advising)

We present Universal Visual Decomposer, an *off-the-shelf* task decomposition method that effectively produces semantically meaning subgoals across both simulated and real-robot environments for long-horizon visual manipulation tasks, without any task-specific knowledge or training. UVD discovered subgoals enable effective reward shaping for solving challenging multi-stage tasks using RL, and policies trained with IL exhibit significantly superior compositional generalization at test time.

We propose a challenging vision-based extrinsic contact localization task: with only a single RGB-D camera view of a robot workspace, identify when and where an object held by the robot contacts the rest of the environment. Our final approach im2contact demonstrates the promise of versatile general-purpose contact perception from vision alone, performing well for localizing various contact types (point, line, or planar; sticking, sliding, or rolling; single or multiple), and even under occlusions in its camera view.

This is the report paper on Workflow Recognition Challenge in MICCAI 2021. I lead the MedAIR team and rank the first over all the 5 rank method in one sub-challenge on multi-modal (videos and kinematics) workflow recognition of robotic surgery videos.

We present a collaborative system with aerial and ground robots to gain precise recognition of feasible region. Taking the aerial robots' advantages of having large scale variance of view points of the same route which the ground robots is on, the collaboration work provides global information of road segmentation for the ground robot, thus enabling it to obtain feasible region and adjust its pose ahead of time.

Inspired by Cerebellum's vital functions in control of human's physical movement, we propose a neural network model of Cerebellum based on spiking neuron networks (SNNs). We apply the model as a feed-forward controller in controlling a 1-DOF robot arm driven by PAMs.

Service

Workshop reviewer: NeurIPS, ICRA
CIS 5190 Applied Machine Learning: TA in Spring 24 at UPenn.
CIS 5200 Machine Learning: TA in Fall 23 at UPenn.
MEAM 5200 Introduction to Robotics: TA in Spring 23 at UPenn.
Fife-Penn Python Club: Instructor in Spring 23 at G.W. Carver High School.

Adwards

GRASP MS Outstanding Research Award 2024
SEAS MS Outstanding Research Award 2024
President Gutmann Leadership Award administered by GAPSA, UPenn, 2023
Best Paper Award at CoRL LEAP workshop, 2023
CoRL 2023 Travel Grant, 2023
Summer Internship Award issued by GAPSA at Penn, 2023
Chiang Chen Overseas Graduate Fellowship (10 students each year in mainland China, 50k$), 2022
Outstanding Graduate of Zhejiang Province issued by Department of Education of Zhejiang Province, 2022
National Scholarship issued by the Ministry of Education of PRC, 2021
First Prize Scholarship issued by Zhejiang University, 2018-2021

Inspired by the template here.