Hung-Chieh Fang

I am a senior undergrad majoring in Computer Science at National Taiwan University, where I am fortunate to be advised by Professors Hsuan-Tien Lin, Yun-Nung (Vivian) Chen and Shao-Hua Sun.

Currently, I am a visiting research intern at Stanford University, working with Prof. Dorsa Sadigh and Amber Xie.

Previously, I was a visiting student at The Chinese University of Hong Kong, where I had the privilege of working with Dr. Yifei Zhang and Prof. Irwin King.

Email  /  CV  /  Google Scholar  /  Github  /  Twitter  /  Linkedin

Research

I am broadly interested in robotics, machine learning and reinforcement learning. My research goal is to develope robots capable of performing complex tasks at or beyond human-level proficiency. I am particularly interested in the following areas:

• Adaptation to distribution shift (e.g., dc-unida, ssd): For robots to be readily used in the real world, they must have almost perfect success rate (> 99%). This requires the ability to adapt (or safely explore) in environments with distribution shifts. I am interested in algorithms that enable test-time adaptation/exploration.
• Dexterous manipulation: To fully unlock the potential of robots in complex tasks such as in-hand manipulation and tool use, fine-grained control is essential. I am interested in how to scale up RL/IL for these tasks.

We explore how to improve representation quality (uniformity) under limited and restricted information due to the nature of federated learning (non-IID, features not shared across clients). We improve inter-client uniformity by softly separating client embeddings without disrupting the underlying data structure.

We found that partial alignment loss fails to outperform the simplest baseline—training only on source data—when there is a source-target label imbalance, due to dimensional collapse in target representations. We address this using de-collapse techniques from self-supervised learning, advancing toward more comprehensive universal domain adaptation.

We propose combining the signal from historical answers with the noise-reduction ability of knowledge distillation to improve information retrieval and question answering.

Projects

We propose a method for retrieving task-relevant data for imitation learning without requiring expert demonstrations. Our approach leverages text descriptions in combination with a vision-language model to enable zero-shot behavior retrieval.

We propose using vision as a surrogate for paired transcripts to enrich the semantic information in self-supervised speech models.

Teaching

This template is adapted from here.