PPO Variants for Humanoid Motion Tracking · Sep. 2025 - Nov. 2025

Xianglei Dong

• Trained and evaluated five PPO variants for motion tracking on a 23-joint Unitree G1 humanoid using NVIDIA Isaac Lab and Isaac Sim.
• Reproduced PPO and Delta Action PPO baselines, then developed an optimized residual-policy variant that decoupled baseline and residual networks to reduce gradient conflicts.
• Achieved a mean reward of 5.03 with a coefficient of variation of 0.068; completed 6,000 iterations in about 2.5 hours on an RTX 4090.

Footage: cropped and enlarged recording of an early Isaac Gym evaluation run. The later experiments described here were conducted in Isaac Lab.

ROS 2 Autonomous Ground Vehicle for IGVC · Team Lead and Core Autonomy Developer, Aug. 2025 - May 2026

Xianglei Dong, Kaisheng Zhang, Boyi Zhang, Boni Zhang

• Led development of the Pioneer 3DX autonomy stack in ROS 2 Humble. I owned mapping and localization with SLAM Toolbox, Nav2 integration and DWB tuning, frontier exploration, TF and motion-control integration, and maintenance of the ROS 2 Pioneer driver.
• Teammates separately owned the E-stop, camera-perception, and LiDAR modules, then integrated the modules and tuned the final task-execution workflow.
• Resolved TF latency, SLAM queue saturation, and lifecycle startup issues. Validation reached -0.077 s TF delay, zero SLAM frame drops, 12.4 cm mean point-navigation error, and a stable 10.5 Hz command loop.

Image: project concept visualization.

Intelligent Home Service Robot Based on ROS · Feb. 2024 - May 2025

Jiawei Xia, Xianglei Dong, Wengcong Zhang, Boyi Zhang, Yutong Sun

• Developed a Kinova Gen2 and Azure Kinect service robot for guest reception, voice interaction, recognition, navigation, cleaning, and manipulation.
• Integrated Azure AI Speech with a finite-state machine; used Cartographer, IRIS LaMa, DWA, and A* for mapping, localization, and planning.
• Connected YOLOv8 3D perception and GraspNet 6D grasp generation to the ROS manipulation stack.

Reproducing and Training ACT from the LeRobot Project · Nov. - Dec. 2024

Jiawei Xia (equal contribution), Xianglei Dong (equal contribution)

• 3D printed, assembled, and brought up a leader-follower robotic arm platform for imitation-learning experiments.
• Reproduced the ACT training pipeline, collected a custom dataset, and trained the arm to organize objects on a desktop.

ICRA RoboMaster University Sim2Real Challenge 2024

Jiawei Xia, Xianglei Dong, Wengcong Zhang, Boyi Zhang

• Estimated box poses from ArUco detections and extracted transformation and rotation matrices.
• Combined an Extended Kalman Filter with an omnidirectional robot to improve localization and pose estimation.

RViz Plugin for Multi-Point Navigation

Xianglei Dong

• Developed a custom RViz tool that records selected goal poses and visualizes them in the navigation interface.
• Implemented automatic multi-point traversal to remove repetitive manual coordinate entry during robot tests.

YOLOv8 GraspNet ROS

Xianglei Dong

• Built a ROS perception pipeline that combines YOLOv8 object detection with GraspNet 6D grasp-pose generation for robotic manipulation.

YOLOv8 3D Detection for ROS

Xianglei Dong

• Developed a ROS Noetic pipeline that combines YOLO detections with depth data to estimate 3D bounding boxes and object poses for navigation and manipulation.