NANYANG TECHNOLOGICAL UNIVERSITY · SINGAPORE南洋理工大学 · 新加坡
Liu Xu刘旭
M.Sc. student in Smart Manufacturing智能制造 · 硕士研究生
My research and engineering work includes SLAM localization integration, 3DGS-based scene reconstruction and object search, WAM-based mobile manipulation, and UAV hardware integration and flight testing.我的研究与工程工作涵盖 SLAM 定位联调、基于 3DGS 的场景重建与物体搜索、基于 WAM 的移动操作，以及无人机硬件集成与实飞验证。
Online 3DGS for object and contextual-anchor search在线 3DGS 中的物体与上下文锚点搜索
Drag right to reveal the real robot向右拖动，查看实机
Simulation仿真
Real robot实机
‹ ›
Simulation evaluates ordered object-and-anchor localization, without grasping or placement; source footage 8×. The real clip shows online scene reconstruction at 4×. These are independent experiments.仿真评估先物体、后锚点的顺序定位，不含抓取或放置；原片 8 倍速。实机片段展示在线场景重建，原片 4 倍速。两段为独立实验。
A past photograph contains both the displaced object and the context of its original location. Online 3D Gaussian Splatting and conditional flow matching guide the robot to locate the object first, then its now-empty anchor.旧照片同时记录了物体和原始位置的上下文。在线 3D Gaussian Splatting 与条件流匹配引导机器人先定位被移动物体，再寻找如今空置的原始锚点。
Key observation主要观察
Restoration requires more than object recognition. Removing contextual-anchor affinity reduces ordered localization success from 34.8% to 8.4% in the reported ablation.复原不仅需要识别物体。在论文消融中，移除上下文锚点关联后，顺序定位成功率由 34.8% 降至 8.4%。
Ordered localization succeeds in 34.8% of held-out simulation episodes, without grasping or placement. In ten real trials, five restorations achieve stable placement on the correct support surface with horizontal error below 0.25 m; final orientation is unconstrained.未见场景中的仿真顺序定位成功率为 34.8%，不包含抓取或放置。十次实机试验中，五次将物体稳定放回正确支撑面，水平误差小于 0.25 m；不约束最终朝向。
Iterative World Action Model for precise scene restoration用于精细场景恢复的迭代世界动作模型
Drag right to reveal the real robot向右拖动，查看实机
Simulation仿真
Real robot实机
‹ ›
Simulation restoration · source footage 4×. Real-robot restoration · source footage 2×, motion smoothed for display. These are independent trials.仿真恢复：原片 4 倍速。实机恢复：原片 2 倍速，含展示用运动平滑。两段为独立试验。
Learned base placement is followed by an iterative World Action Model that refines future predictions before generating arm actions. The robot executes a short action segment, observes the scene again and replans.学习式底盘停靠之后，迭代世界动作模型在生成机械臂动作前细化未来预测。机器人执行一段预测动作，再获取新观测并重新规划。
Key observation主要观察
A feasible docking pose need not be useful for restoration. The pose utility is supervised by outcomes from a frozen manipulation model; in the three-training-seed simulation ablation, removing this utility predictor reduces mean success from 71.1% to 47.3%.可行的停靠位姿不一定有利于恢复。停靠效用以冻结操作模型的恢复结果为监督；在三个训练种子的仿真消融中，移除效用预测器后，平均成功率由 71.1% 降至 47.3%。
19/30 successful real trials at a 30 mm tolerance. The mean final position error is 30.7 ± 12.7 mm across all 30 trials, including failures.在 30 mm 容差下，实机成功 19/30 次。全部 30 次试验（含失败）的平均终态位置误差为 30.7 ± 12.7 mm。
Contributed to LIO/FAST-LIO debugging: scan-to-map matching, initial pose alignment and odometry anomaly investigation. Contributed to initial scan alignment using Manhattan-world features and IMU gravity to address initial alignment error and height drift, and to relocalization in the official coordinate frame.负责部分 LIO/FAST-LIO 定位调试：scan-to-map 匹配、初始位姿对齐与 odom 异常排查；参与结合 Manhattan 世界特征与 IMU 重力估计进行初始扫描对齐，以处理初始对齐误差和高度漂移，并参与官方坐标系中的重定位联调。
10 Hz → 100 HzContributed to increasing the LIO update frequency through debugging and integration.参与调试与集成，提升 LIO 更新频率。
ONBOARD INTEGRATION机载系统集成
UAV integration无人机硬件集成
Participated in selection and integration of Livox Mid-360 LiDAR, monocular cameras, Jetson Orin NX, flight controllers and airframes. Handled part of the mounting, wiring, power connections, interface configuration and troubleshooting for three UAVs.参与 Livox Mid-360、单目相机、Jetson Orin NX、飞控及机架的选型和集成，负责部分安装固定、接线与供电连接、接口配置、联调及故障排查，支持 3 架无人机飞行验证。
The integration spans sensor inputs, localization output, planner trajectories and the flight-controller interface. Simulation used ROS2 and Unity HiFi Simulator. The three-UAV deployment from December 2025 to May 2026 used ROS1 on Ubuntu 20.04; I contributed to localization debugging and onboard integration.集成涉及传感器输入、定位输出、规划轨迹与飞控接口。仿真阶段使用 ROS2 与 Unity HiFi Simulator；2025.12—2026.05 的三机实飞阶段使用 ROS1／Ubuntu 20.04，我参与定位联调和机载系统集成。
Team validation团队验证
<10 cmLocalization error in module tests模块测试定位误差
<10 cm @ 3 m/sTrajectory tracking error动态轨迹跟踪误差
3/3 · 72 sDemo Day · UAVs completedDemo Day · 完成无人机数
Basic Sequence / Alternate Sequence / Demo Day: 50 / 60 / 72 s, each with 3/3 UAVs completing the task. Overall real-world completion-speed rank: 2nd of 8 teams.Basic Sequence / Alternate Sequence / Demo Day 分别为 50 / 60 / 72 s，每项均为 3/3 架完成；实机验收整体完成速度排名 2/8。
Simulation contribution仿真阶段分工
From July to November 2025, I led EGO-Planner adaptation for local planning, obstacle avoidance and replanning in the ten-UAV simulation. I adjusted safety margins, velocity constraints, takeoff order and task flow for narrow gates, corners and inter-agent path conflicts, and integrated ROS2 interfaces for mission initialization, odometry, cameras, trajectories and results. The team completed 10/10 UAVs in 166 s, ranking 1st of 8 teams by completion speed.2025.07—2025.11，我负责十机仿真中的 EGO-Planner 局部规划、避障与重规划适配，针对窄门、转角和多机路径冲突调整安全距离、速度约束、起飞顺序与任务流程，并接入 ROS2 的任务初始化、里程计、相机、轨迹和结果接口。团队仿真验收 10/10 架完成，用时 166 s，完成速度排名 1/8。
03
What the experiments show实验说明了什么
In Spot the Difference's ten real trials, the main observed failure stage is locating the original position. For Copy-Paste, a third refinement adds only 0.2 percentage points of mean simulation success and increases planning-round latency on RTX 6000 Ada from 0.654 to 0.838 s.Spot the Difference 的十次实机试验中，主要失败环节是原始位置搜索。Copy-Paste 增加第三轮预测细化后，仿真平均成功率只提高 0.2 个百分点，而 RTX 6000 Ada 上的规划轮时延从 0.654 s 增至 0.838 s。
04
Background教育与经历
Education教育背景
2025.08 — Present至今
Nanyang Technological University南洋理工大学
M.Sc. in Smart Manufacturing · Singapore智能制造 · 硕士研究生 · 新加坡
Supported line maintenance and fault analysis; used Python and Excel to structure, classify and analyze maintenance records.参与航线维修及故障排查，使用 Python 与 Excel 结构化维修记录，分析异常类别和重复性故障。
2023.07 — 2023.08
Bofeike (Tianjin) Technology博菲克（天津）科技有限公司
Mechanical Engineer · R&D Department机械工程师 · 研发部
Worked on automotive fluid-pipeline component design, UGNX modeling and assembly checks, and ANSYS structural analysis.参与汽车流体管道部件设计，使用 UGNX 建模与装配检查，并基于 ANSYS 开展结构分析。
OTHER MANUSCRIPT / THIRD AUTHOR其他研究稿件 / 第三作者
Uncertainty-Aware Dynamics Learning via Offline Generative Prior Extraction and Online Bayesian Posterior Refinement
IEEE Transactions on Industrial Informatics · Regular Paper · Under review · April 2026Regular Paper · 审稿中 · 2026 年 4 月
