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, Shao-Hua Sun and Yun-Nung (Vivian) Chen.

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

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.

I am applying for a PhD position in 2026 Fall.

Email  /  Google Scholar  /  Github  /  Twitter  /  Linkedin

News

Research

I am broadly interested in robotics, representation learning and reinforcement learning. My goal is to understand and develop robots with generalist intelligence that can autonomously adapt in the real world and ultimately perform tasks in our daily lives. I am particularly interested in:

• Representation learning for generalist policies: (1) learning generalized representations from heterogeneous data (2) discovering the underlying shared structure across different modalities (e.g., vision, text, control, touch).
• Representation learning for adaptation: (1) enabling robust adaptation to distribution shift (2) developing unsupervised reinforcement learning algorithms for rapid online exploration.
• Dexterous manipulation: scaling up dexterous hand capabilities across diverse tasks by learning from diverse sources (e.g., simulation, human videos).

We explore how to improve generalization under highly non-IID client distributions where representations are non-shared. We propose soft separation, which regularizes each client toward a distinct subspace to improve inter-client uniformity (maximizing information) while preserving semantic alignment.

We study how to adapt to arbitrary target domains without assuming any class-set priors. Existing methods suffer from severe negative transfer under large class-set shifts due to the overestimation of importance weights. We propose a simple uniformity loss that increases the entropy of target representations and improves performance across all class-set priors.

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

Course Projects

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

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.

Teaching

This template is adapted from here.