Jianshu Hu

I'm a PhD student at UM-SJTU Joint Institute in Shanghai, working on sample-efficient and generalizable robot learning. I am currently advised by Yutong Ban and Paul Weng. Before my PhD study, I finished my master degree in Upenn GRASP, advised by Michael Posa.

Email  /  Scholar  /  Github

profile photo

Research

I'm interested in reinforcement learning and robot manipulation. My current research focuses on improving sample-efficiency and generalization ability of robot learning algorithms by exploiting data augmentation, leveraging pre-trained models, learning a dynamics model/world model. Some papers are highlighted.

b3do State-Novelty Guided Action Persistence in Deep Reinforcement Learning
Jianshu Hu, Paul Weng, Yutong Ban,
Machine Learning Journal
arXiv

In this paper, we propose a novel method to dynamically adjust the action persistence based on the current exploration status of the state space.

b3do Revisiting Data Augmentation in Deep Reinforcement Learning
Jianshu Hu, Yunpeng Jiang, Paul Weng,
ICLR, 2024
arXiv

We make recommendations on how to exploit data augmentation in image-based DRL in a more principled way. And we include a novel regularization term called tangent prop in RL training.

b3do Solving Complex Manipulation Tasks with Model-Assisted Model-Free Reinforcement Learning
Jianshu Hu, Paul Weng,
CoRL, 2022
paper link

A novel deep reinforcement learning approach for improving the sample efficiency of a model-free actor-critic method by using a learned model to encourage exploration.

b3do Beyond Inverted Pendulums: Task-Optimal Simple Models of Legged Locomotion
Yu-Ming Chen, Jianshu Hu, Michael Posa,,
T-RO
arXiv

We propose a model optimization algorithm that automatically synthesizes reduced-order models.

Miscellanea

Reviewer, ICLR 2024
Reviewer, TNNLS

This website is build on this source code. Thanks to Jon Barron.