Tong Xu

PhD student for Robotics

Short Bio

I am Tong Xu (徐通), a PhD student at George Mason University, advised by Prof. Xuesu Xiao. I hold a M.S. Degree from University of Southern California and a B.E. Degree from Nanjing University of Information Science & Technology.

My primary research interests include robotics, reinforcement learning and off-road navigation. My current work on automatic curriculum learning for field and legged robotics reflects my dedication to advancing real-world reinforcement learning applications in robotics through cutting-edge AI techniques.

Interests

Reinforcement Learning
Field Robotics
Legged Robotics

Education

PhD in Computer Science, 2023
George Mason University
M.S. in Computer Science, 2021
University of Southern California
B.E. in Network Engineering, 2017
Nanjing University of Information Science & Technology

News

09/2024: One paper about reinforcement learning navigation is accepted by 2024 IEEE International Symposium on Safety, Security, and Rescue Robotics (SSRR)!
08/2023: I began my PhD journey in RobotiXX Lab at George Mason University, supervised by Prof. Xuesu Xiao.
05/2023: I graduated from University of Southern California.
05/2022: I joined Human-to-Everything (H2X) Lab as visiting researcher at Boston University, supervised by Prof. Eshed Ohn-Bar.
06/2021: I graduated from Nanjing University of Information Science & Technology.

Experience

Graduate Research Assistant

RobotiXX

May 2024 – Aug 2024 Fairfax

Automatic Curriculum Learning (ACL) for Verti-Wheelers

Graduate Teaching Assistant

George Mason University

Aug 2023 – Present Fairfax

Designed student lab contents involving data structure and leaded the whole lab recitation
Created grading scripts and managed a team of 6 undergraduate teaching assistants

Visiting Researcher

H2X Lab

May 2022 – Aug 2022 Boston

DeepVO - Visual Odometry with Deep Learning
OpenGuide - A Scalable Human-Like Guidance System for Travelers with Visual Impairment

Recent Publications

Quickly discover relevant content by filtering publications.

Verti-Selector: Automatic Curriculum Learning for Wheeled Mobility on Vertically Challenging Terrain

Verti-Selector, an automatic curriculum learning framework designed to enhance learning efficiency and generalization by selectively sampling training terrain

Tong Xu, Chenhui Pan, Xuesu Xiao

PDF Video

Verti-Selector: Automatic Curriculum Learning for Wheeled Mobility on Vertically Challenging Terrain

Pietra: Physics-informed evidential learning for traversing out-of-distribution terrain

Physics-Informed Evidential Traversability (PIETRA), a self-supervised learning framework that integrates physics priors directly into the mathematical formulation of evidential neural networks and introduces physics knowledge implicitly through an uncertainty-aware, physics-informed training loss

Xiaoyi Cai, James Queeney, Tong Xu, Aniket Datar, Chenhui Pan, et al.

PDF Video

Pietra: Physics-informed evidential learning for traversing out-of-distribution terrain

Reinforcement learning for wheeled mobility on vertically challenging terrain

A custom-designed simulator (VW-Chrono) built on the Chrono multi-physics engine, our approach leverages Proximal Policy Optimization (PPO) and a terrain difficulty curriculum to refine a policy based on a reward function to encourage progress towards the goal and penalize excessive roll and pitch angles, which circumvents the need of complex and expensive kinodynamic modeling, planning, and control.

Tong Xu, Chenhui Pan, Xuesu Xiao

PDF

Reinforcement learning for wheeled mobility on vertically challenging terrain

See all publications

Projects

Automatic Curriculum Learning for Verti-Wheelers

Proposed Verti-Selector, a novel ACL framework that samples training terrain based on estimates of future learning potential (TD-error). Significantly enhanced navigation performance by 23.08% improvement in terms of success rate compared against a manually designed curriculum, vanilla Reinforcement Learning, and two baseline approaches

Paper Video

DeepVO

Reproduced DeepVO network architectures by integrating FlowNetSimple pretrained model with LSTM. Improved the influence of epochs by 14.8% and hyperparameter by 8.4% in the loss function of DeepVO network, using KITTI and nuScenes datasets, based on translation and rotation RMSE values

Code Presentation

Cone Detection

Reproduced Faster R-CNN model by integrating Region Proposal Network module with Fast R-CNN. Improved the performance of Faster R-CNN model by 37.4% in terms of recall rate compared with YOLOv3 model on cone annotated dataset

Code Presentation Report

OpenGuide

A Scalable Human-Like Guidance System for Travelers with Visual Impairment.

Code Website