Jianming Xing<br> 邢鉴明

Jianming Xing
邢鉴明

Second B.S. @ HIT Shenzhen

🎓 Education

  • 2025.09 - 2027.06, Second Bachelor’s Degree in Computer Science and Technology, Harbin Institute of Technology, Shenzhen.
  • 2021.09 - 2025.06, B.Eng. in Mechanical Design, Manufacturing and Automation, Harbin Institute of Technology, Weihai.

🤖 Projects

1) VLA/OpenPI Real-Robot Data Collection and Inference Loop for CR5 — ROS2 Humble · Dobot CR5 · LeRobot v2.0 · OpenPI
Role: Hand-eye calibration, teleoperation, synchronized data collection, safety validation, OpenPI adaptation, and real-robot inference deployment · Period: 2026.01 - 2026.03
Built a real-robot closed loop for Dobot CR5, Orbbec Astra2 RGB-D, and an electric gripper, covering calibration, teleoperation, synchronized data collection, dataset conversion, OpenPI fine-tuning, and WebSocket-based inference.
  • Designed an HDF5 synchronization pipeline using RGB image timestamps as the master clock, aligning robot state, target action, and gripper feedback with binary search and linear interpolation.
  • Reduced early camera frame drops from 27.15% to 2.57%-4.74% by using Fast DDS shared memory and RGB-only capture.
  • Completed OpenPI adaptation with Dobot 7-D state/action mapping, training configuration, WebSocket policy serving, and a CR5 inference client using action chunks and blocking short-horizon execution.
Outcome: Established a safety-checked real-machine VLA pipeline from data recording to LeRobot v2.0 conversion and CR5 inference deployment.
Embodied AI VLA ROS2 OpenPI LeRobot Real Robot
2) Robot Digital-Twin HMI for Weld-Seam Recognition — C++/Qt upper computer · U-Net seam segmentation · OpenGL digital twin
Role: Vision model, upper-computer control, and digital twin integration · Period: 2024.09 - 2025.06
Built a robotics software stack for weld-seam recognition and LeArm digital-twin demos, combining semantic segmentation, Qt-based control, and real-time OpenGL visualization.
  • Compared baseline U-Net, VGG-U-Net, and a VGG16-pretrained variant, improving weld-seam segmentation accuracy from 64% to 96.8% through transfer learning, data augmentation, and dataset expansion.
  • Integrated image selection, Python inference calls, and segmentation-result display into a C++/Qt HMI.
  • Implemented OpenGL-based digital twin rendering with STL link models, DH-parameter kinematics, slider control, and mouse interaction.
Outcome: Integrated weld-seam detection and robot visualization into one desktop workflow, with resume-backed segmentation accuracy reaching 96.8%.
Robotics Computer Vision C++/Qt U-Net OpenGL

💻 Experience

Xbotics Embodied AI Community Internship
MotrixLab, Isaac Lab, quadruped robotics, reinforcement learning
2025.10 – 2026.01
  • Migrated Isaac Lab's ANYmal-C navigation task to a MotrixLab NumPy environment, rebuilding reset/step, command sampling, observation assembly, reward calculation, and termination checks while preserving the 12-D joint-position action space, 54-D policy observation, and goal position/yaw navigation interface.
  • Implemented MuJoCo Heightfield terrain height/slope queries with Y-axis frame correction, spawn/goal filtering by elevation and slope, spawn Z alignment, and boundary termination; added vertical-velocity, body-attitude, and foot-contact rewards plus registry, RL config, zero-action rollout, and reward/termination regression tests.
HERO Robomaster Team
Vision Framework and ROS2 Migration Developer · Harbin Institute of Technology, Weihai
2023.09 – 2024.02
  • Migrated and modularized the Robomaster vision framework to ROS2 nodes, replacing native C++ threading to improve real-time image acquisition, maintainability, and development scalability.
  • Handled MCU-to-Linux host receiving, ROS2 node message exchange, and time synchronization with the MCU.
  • Adjusted the communication protocol for new competition-season requirements and improved the foundation for later perception and aiming modules.