Wenjing Margaret Mao
[毛文婧]

I'm a second-year Robotics MSE student at the GRASP Lab, University of Pennsylvania, where I am honored to work with Prof. Antonio Loquercio.

Prior to this, I received my bachelor's degree from NYU Shanghai, double majoring in Computer Science and Data Science(AI track), and graduated with honors in Computer Science. During my undergraduate studies, I was honored to work with Prof. Lerrel Pinto at the General-Purpose Robotics and AI Lab(GRAIL) within the CILVR Lab at NYU and Prof. Gus Xia at MusicXLab, who supported my broad exploration of both robotics and computer music. I also worked as a Machine Learning Research Intern on the IPEX-LLM team at the Intel Asia-Pacific Research and Development Center in Shanghai.

My research focuses on how robots learn from multi-sensory inputs, including vision, touch, sound, and proprioception, to understand and interact with the environment. I aim to develop intelligent systems that bridge human and robotic intelligence by translating rich sensory cues into embodied motor behaviors. Broadly, I explore how cross-modal perception and human-inspired sensory learning can enable robots to reason, adapt, and act seamlessly in dynamic, real-world environments.

Email  /  GitHub  /  LinkedIn

Research

RoSHI: A Versatile Robot-oriented Suit for Human Data In-the-Wild

RoSHI is a hybrid motion-capturing system that enables large-scale, in-the-wild human-to-robot data collection, supporting not only motion imitation but also the training of robot behaviors from human data.

Audio Perception in Robot Manipulation

This project explores how principles of movement sonification (the mapping of motion to sound) can inspire more perceptive and adaptive robotic systems. By analyzing how auditory cues convey contact forces during manipulation, I examined how sound can complement visual feedback in robot learning. The study highlights the potential of integrating audio perception into multimodal robot representations, bridging insights from human motor cognition and embodied AI.

Selected Projects

ParkourBot: Adaptive Locomotion & Obstacle Navigation via Hierarchical Reinforcement Learning

ParkourBot extends the Unitree RL Gym with a two-tier hierarchical reinforcement learning pipeline that enables a Unitree G1 quadruped to walk and high-step through dense, obstacle-rich mazes. A high-level Q-learning planner reasons over a 3-D grid map to coordinate specialised low-level PPO gait controllers, producing parkour-style movements in simulation.

Dynamic Pick-and-Place Motion Planning with Franka Arm

In this project, our system autonomously performs a dynamic pick-and-place-and-stack task for both static and moving objects. By integrating vision and advanced motion planning, we achieve high-precision manipulation capable of stacking cubes on both stationary and rotating platforms.Dynamic pick-and-place motion planning for robotic manipulation using the Franka arm. This project focuses on real-time motion planning and control for complex manipulation tasks.

Autonomous F1TENTH Car Racing

Developed three stages of an autonomous F1TENTH race car using Follow-the-Gap, Pure Pursuit, and MPPI algorithms, ranging from closed-track racing without racing lines to extreme racing-line following and head-to-head competition on complex outdoor tracks. Achieved real-time racing and obstacle avoidance through full perception–planning–control integration in ROS 2, enabling dynamic lane switching and overtaking.

Sokoban Bot: An LLM Agent for Object Pushing Navigation

Developed an LLM-grounded navigation agent integrating reasoning-based path planning and feedback correction for object- pushing tasks, enabling the F1TENTH car to interpret grid maps, plan obstacle-free trajectories, and iteratively refine actions through error-aware memory in ROS 2.

Design and source code from Jon Barron's website