Zheyuan Hu

I am an incoming Ph.D. student at CMU Robotics Institute. I received my Bachelor’s degrees in Computer Science and Applied Mathematics from UC Berkeley, advised by Professor Sergey Levine. Previously, I had the privilege of working with Professor Abhishek Gupta and Dr. Kelvin Xu.

My research interests are in robot learning, specifically in methods enabling robots to acquire skills rapidly, autonomously, and scalably in the real world. This includes sample-efficient reinforcement learning, learning from diverse data, reset-free and continual learning. Recently, I have worked on dexterous manipulation challenges, reset-free learning, solving long horizon tasks, and reusing prior data to accelerate acquisition of new tasks in real world settings.

My next research goal is to build algorithms and systems that allows robots to learn manipulation skills under minutes of real world interactions.

GitHub  /  Google Scholar  /  LinkedIn

Research

Yell At Your Robot: Improving On-the-Fly from Language Corrections

YAY Robot leverages verbal corrections to enable on-the-fly adaptation and continuous policy improvement on complex long-horizon tasks.

SERL: A Software Suite for Sample-Efficient Robotic Reinforcement Learning

SERL (Sample-Efficient Robotic reinforcement Learning) provides an open-source software framework that aims to facilitate wider adoption of RL in real-world robotics.

REBOOT: Reuse Data for Bootstrapping Efficient Real-World Dexterous Manipulation

REBOOT learns dexterous manipulation skills autonomously and entirely in the real world in less than 8 hours by bootstrapping from prior data. The method achieves 2X speed up than learning from scratch. Our method is tested on a multi-fingered robot hand learning dexterous in-hand rotation tasks.

Dexterous Manipulation from Images: Autonomous Real-World RL via Substep Guidance

AVAIL allows robots to learn long horizon manipulation skills through autonomous real-world interaction without manual engineering, using a framework for users to breakdown a long task into multiple sub-tasks with image examples and deep RL. The method was tested on a four-finger robot hand and complex dexterous manipulation tasks in the real world.

Design and source code from Jon Barron's website