许广云许

Embodied AI Algorithm Engineer | Robot Manipulation · VLA · 3D Vision

About

I am a robotics and embodied AI engineer working at the intersection of robot perception, 3D vision, and intelligent manipulation.Currently I am a Founding Engineer at FiveAges, where I design and deploy embodied AI systems for real-world robotic platforms including dual‑arm robots and humanoid robots. My work focuses on vision‑guided manipulation, perception pipelines for robot operation, and deploying Vision‑Language‑Action (VLA) models to enable robots to execute complex tasks from natural language instructions.Previously I worked as a perception algorithm engineer at DeepGlint and Zhixingzhe Technology. My work covered 3D reconstruction, point cloud processing, LiDAR perception, and multi‑object tracking for autonomous systems.At DeepGlint, I developed point‑cloud‑based anomaly detection algorithms and achieved state‑of‑the‑art performance on the MVTec 3D‑AD benchmark. I also built video‑based 3D reconstruction pipelines using COLMAP, hloc, NeRF, and Gaussian Splatting.At Zhixingzhe Technology, I worked on LiDAR perception for autonomous driving, including 3D object detection, dynamic/static detection, and multi‑object tracking. Improvements to the tracking pipeline increased the MOTA metric by more than 20 percentage points while maintaining real‑time performance.I received my M.S. in Computer Technology from the University of Chinese Academy of Sciences (CASIA). My research focused on robot perception and grasping, resulting in publications at ICRA and Frontiers in Neurorobotics.Research InterestsEmbodied AI · Robot Manipulation · 3D Vision · Point Cloud Processing · Robot Perception · Vision‑Language‑Action Models

Country

China

City

Haidian District

Industry

Internet

Skill

Experience

Founding Engineer

FiveAges

2025-3 - Present · 1 yr 7 mos

海淀区

• Designed and deployed vision-guided manipulation systems for dual-arm and humanoid robots, enabling reliable object picking and placement in real-world scenarios • Developed and demonstrated embodied AI applications including autonomous microwave manipulation and unstructured object handling • Deployed Vision-Language-Action (VLA) models on real robotic hardware, bridging language instructions with physical robot actions • Built full-stack perception pipelines covering hand-eye calibration, depth estimation, and 3D vision for manipulation

Research And Development Engineer

DeepGlint

2023-5 - 2025-3 · 1 yr 11 mos

北京市, 中国

• Achieved SOTA on MVTec 3D-AD benchmark: +1.0% detection accuracy and +0.9% segmentation accuracy over the prior #1, using a custom point cloud feature extractor combined with sparse coding and memory bank • Built video-based 3D reconstruction pipelines using COLMAP, hloc, and NerfStudio; improved reconstruction accuracy with temporal sequential matching and adaptive frame sampling based on optical flow • Developed industrial point cloud measurement system for part dimension inspection (e.g., brake pad thickness, pantograph wear) using point cloud registration — deployed in real manufacturing environments • Researched and benchmarked state-of-the-art point cloud registration methods (FCGF, GeoTransformer, BUFFER, MAC); identified FCGF and GeoTransformer as best for industrial plug-and-play deployment

Research And Development Engineer

Beijing IDRIVERPLUS Technology Co., Ltd.

2021-7 - 2023-5 · 1 yr 11 mos

海淀区

• Improved LiDAR multi-object tracking algorithm: achieved 20+ percentage point improvement in MOTA over Apollo baseline, inspired by ImmortalTracker, with inference time under 10ms on Xavier platform • Designed heading angle filtering for tracked objects (static / straight-line / turning state classification), achieving ~2 mAP improvement in heading estimation • Developed LiDAR dynamic/static detection using BEV-based optical flow estimation (MotionNet / PillarMotion framework), adapted for Horizon BPU deployment • Built secondary clustering module for missed-detection recovery using BEV projection, OpenCV, and PCL