I am a first-year master student in the Robotics Institute at Carnegie Mellon University, working with Prof.Fernando De La Torre and Dr.Joseph K J on evaluation for vision language models.

During my undergrad, I worked with Prof.Huazhe Xu at Tsinghua University on reinforcement learning. I have also collaborated with Dr.Ruohan Zhang at Prof.Jiajun Wu’s Lab at Stanford University on Human-robot interaction. I also spent a summer doing internship at Microsoft WebXT group, Bing Ads team.

News

• One paper accepted at Neurips’23
• I joined Carnegie Mellon Univerisy for the Master’s in Robotics (MSR) program!
• I graduated from the Computer Science Department of Tsinghua University with a Bachelor’s degree in engineering!
• Our paper “A Dual Representation Framework for Robot Learning with Human Guidance” is accepted by CORL’22
• I gave a talk at Stanford Vision & Learning Lab
• I became a visiting research intern at Stanford in July, 2022
• I became a Deep Learning Intern at Microsoft China in July, 2021
• Our paper “The Origin of CBRAM With High Linearity, ON/OFF Ratio, and State Number for Neuromorphic Computing” is accpeted at IEEE-TED, 2021

Projects

• Robot Learning with Text-to-Image Generation Models
  • Explore the possibility of utilizing the knowledge embedded in pre-trained text2image diffusion models to provide visual guidance for robot learning, without any expert demonstrations or human-designed reward functions.
  • Propose a pipeline using image editing techniques to generate visual goals for example-based visual RL, and experiments in simulation and real-world validates the effectiveness of the proposed method.
• Dual Representation for Robot Learning under Human Guidance
  • Focus on designing sample-efficient preference learning algorithms that can achieve high performance in both oracle and human experiments.
  • Incorporate scene graph as a high-level abstraction representation to facilitate better query selection algorithms in preference learning, which lead to SOTA performance in reward estimation.
• Concurrent Multi-task Learning
  • Come up with a new task setting, “concurrent multi-task learning”, where the demonstrator demonstrates multiple tasks in a single demonstration, and the algorithm tries to learn all separate tasks with these “concurrent” demonstrations.
  • Propose a EM-based Neural Network learning algorithm that can effectively learns separate tasks without requiring extra labels.
• Neuromorphic Computing
  • Explore the origin of CBRAM with High Linearity, ON/OFF Ratio, and State Number, which are crucial for a successful neuromorphic computation network(such as deep-learning neural networks) built from these CBRAMs as neurons.
  • Build a Powerful Neural Network using this CBRAM model as neurons which can solve the MNIST classification problem with SOTA performance(92%) compared to previous neuromorphic computation networks(80%).