Yunzhi Zhang

I am a first year Ph.D. student in computer science at Stanford University.
Previously, I received B.A. in computer science and pure mathematics from UC Berkeley. I worked on reinforcement learning, advised by Pieter Abbeel and Lerrel Pinto.

Scholar  /  Github  /  Email  /  LinkedIn

Automatic Curriculum Learning through Value Disagreement
Yunzhi Zhang, Pieter Abbeel, Lerrel Pinto
[arXiv] [Webpage] [Code]
NeurIPS 2020.

We introduce a goal sampling strategy for goal-conditioned robotics tasks, where goals lying on the learning frontier are more likely to be selected for training. Such goals are identified as goals with high Q-function uncertainty. This simple technique is able to provide a strong learning signal even in sparse reward environments.

Asynchronous Methods for Model-based Reinforcement Learning
Yunzhi Zhang*, Ignasi Clavera*, Boren Tsai, Pieter Abbeel
[arXiv] [Webpage] [Code]
CoRL 2019 (Spotlight).

We propose a general framework for model-based reinforcement learning methods with asynchronous data collection, dynamics model training and policy learning. It achieves wall-clock-time-wise efficiency and reduces overfitting on several complex robot manipulation tasks in the real world.


Honors and Awards

Design from Jon Barron's website.