CV

Non-Verbal Human Behavior Understanding for Interactive VLA Systems

University of Wisconsin-Madison | NSF HAND ERC

• Prototyped a hierarchical VLM-VLA system for non-verbal HRI that maps visual observations of human behaviors and human-object interactions to context-aware robot instructions.
• Integrated fast and reasoning Qwen VLMs with π0.5 execution, enabling high-level behavior interpretation and planning to drive vision-conditioned low-level robot actions.

Bimanual Dexterous Manipulation for Real-World Jar Opening

University of Wisconsin-Madison | NSF HAND ERC

• Developed a unified PPO policy for bimanual dexterous manipulation, learning grasp acquisition, grasp-to-twist transition, and sustained lid twisting without explicit task-phase input; achieved 88% success (53/60 trials) across six real-world household containers.
• Introduced a geometric hand-object enclosure representation with a continuous measure to guide grasp formation and a binary indicator to encode the grasp-to-twist transition within a unified policy.
• Designed a three-stage curriculum for contact-reconfiguration exploration and sim-to-real adaptation, enabling emergent finger gaiting to sustain lid rotation beyond the motion range of a fixed grasp.
• Scaled asymmetric PPO to 4,096 GPU-parallel environments in mjlab/MuJoCo and deployed the policy zero-shot on dual Franka arms with TESOLLO dexterous hands and RGB-D object state estimation.

Immersive Robotic Telepresence via Online 3D Gaussian Splatting

University of Wisconsin-Madison

• Engineered and deployed a real-world VR robotic telepresence system integrating Stretch 3, RealSense RGB-D sensing, Meta Quest/OpenXR, and a remote online 3D reconstruction and rendering pipeline.
• Implemented a ZeroMQ-based RGB-D streaming pipeline for external RealSense cameras, bypassing ROS 2 image topics for low-latency camera transport between the robot and remote host.
• Integrated OpenXR/Meta Quest teleoperation with the Stretch ROS 2 stack for bidirectional robot control and immersive live remote exploration.

Real-Time Social Navigation in Human Environments

University of Michigan-Ann Arbor | Fluent Robotics Lab

• Developed a real-time social navigation system and research platform on a Hello Robot Stretch, extending its ROS stack for data collection.
• Built a real-time perception pipeline with YOLOv8 to detect and track human positions and skeletal keypoints from an RGB-D camera, using a Kalman filter variant to estimate 3D features from point-cloud depth.
• Implemented Human Scene Transformer to forecast pedestrian trajectories and model predictive control to avoid collisions with pedestrians.

Amphibious Quadruped Robot with Reinforcement Learning

Westlake University | I4FSI Lab

• Developed a 12-DoF amphibious quadruped robot, designing thrust and buoyancy modules for underwater adaptation and a Raspberry Pi 4B control platform with servo and IMU interfaces.
• Built an underwater inverse-kinematics model and an empirical LSTM thrust model.
• Trained PPO policies in a custom Gym environment for trajectory control and designed long-duration underwater experiments for reinforcement learning.

Electronics Engineering Intern

ABB

Ph.D.

University of Wisconsin-Madison

Computer Sciences

M.Eng.

University of Michigan-Ann Arbor

Electrical and Computer Engineering

B.Eng.

Dalian University of Technology

Electrical Engineering

Grasp2Twist: Learning Bimanual Dexterous Jar Opening by Reinforcement Learning

Toward Real-time High-fidelity Immersive Robotic Telepresence via Online 3D Gaussian Splatting