linji · robots that learn, one task at a time
↩ OS

Selected robot learning research map for Linji Wang

Linji Wang
WHOAMI

Linji (Joey) Wang

AI robotics + systems engineer and CS Ph.D. researcher @ George Mason University · RobotiXX Lab, advised by Xuesu Xiao. Automatic curriculum learning × deep RL × embodied agents.

robots that learn, one task at a time
TIMELINE

The path so far

2023 — now
Ph.D. Computer Science, George Mason
RobotiXX Lab · automatic curricula and adaptive rewards
2026 summer
SDE Intern, AWS (Aurora PostgreSQL)
C · database internals · Adaptive Join
2025 summer
SDE Intern, AWS (RDS Proxy)
statistical performance-regression testing
2021 — 2023
M.S. Mechanical Engineering, CMU
3D perception · AR scene inpainting · GPA 3.94
2016 — 2021
B.S. Mechanical Engineering, U. Cincinnati
where the robots started
RESEARCH THESIS

Make training adapt to the learner

Estimate capability online, then change tasks or auxiliary rewards so the robot practices the right challenge at the right time.

observe→estimate→adapt→act
IROS 2025 · FIRST AUTHOR

GACL: Grounded Adaptive Curriculum Learning

GACL framework

A teacher generates training tasks from online learner performance, grounded in limited reference samples from the target task distribution. With Zifan Xu, Peter Stone, Xuesu Xiao.

  • +6.8% success · wheeled navigation
  • +6.1% success · quadruped locomotion
arXiv project
IROS 2025 · CO-FIRST AUTHOR

Reward Training Wheels

RTW framework

Auxiliary rewards that adapt as the robot learns — training wheels that fade away. With Tong Xu, Yuanjie Lu, Xuesu Xiao.

  • simulation: +122.62% mobility · 3× faster to threshold
  • physical: 5/5 trials vs 2/5 baseline
arXiv project
IROS 2025 · 🏆 2ND PLACE, BARN 2025 (SIM + REAL)

Decremental Dynamics Planning

Full robot dynamics near the robot, progressively simpler along the horizon. Led by Yuanjie Lu; I'm 3rd author, with Tong Xu, Nick Hawes, and Xuesu Xiao.

  • improved 3 planners overall
  • 2nd in both simulation and physical phases
arXiv project
IEEE RA-L 2025

II-NVM: Normal Vector-Assisted Mapping

II-NVM system overview

Surface-normal consistency fixes double-sided mapping of thin structures in SLAM. Led by Chengwei Zhao; I'm 5th of 7 authors.

arXiv code project
TOOLBOX

Skills & stack

Curriculum Learning Reward Design PPO Isaac Gym / Lab MuJoCo ROS / ROS 2 ONNX Runtime PyTorch C C++ PostgreSQL Query Optimization AWS
HARDWARE

Robot platforms

🛞 wheeled UGV — constrained navigation 🐕 quadruped — confined 3D spaces 🏔️ off-road vehicle — vertical terrain 🤖 humanoid — MTC 2026

🏆 2nd in both phases

2025 BARN Challenge — RobotiXX DDP system, simulation qualifier + physical finals

ARXIV 2026 · HUMANOID LOCOMOTION

Moving Through Clutter

VR data collection and benchmarking for scene-aware humanoid locomotion in dense 3D clutter. I am 3rd author, with Beichen Wang, Yuanjie Lu, Liuchuan Yu, and Xuesu Xiao.

  • 348 whole-body trajectories
  • 145 diverse cluttered scenes
arXiv project
// selected research · a linear version of the guided canvas tour
100%
wheel zoom · drag pan · tour guided camera