Liu Xu刘旭

From a photograph to the current 3D scene从目标照片到当前三维空间

A past photograph records an object and its relationship to the surrounding scene. Once the object moves, restoration requires finding both the displaced object and its now-empty original location. Object appearance helps with the first search; recognizing the second also requires the surrounding context.旧照片记录了物体及其与周围环境的关系。物体被挪走后，复原需要同时找到被移动的物体和已经空置的原位置。物体外观有助于搜索前者，识别后者还需要利用周围环境的上下文。

Online 3DGS integrates camera observations as the robot explores. The current map and goal photograph condition a flow-matching model that produces an eight-channel Restoration Field for object, anchor and exploration readout. Persistent readout in observed regions connects spatial predictions to Nav2 navigation and Contact-GraspNet grasp proposals, linking reconstruction, search and physical execution.机器人探索时，在线 3DGS 持续整合相机观测。当前地图与目标照片共同驱动条件流匹配模型，生成八通道恢复场，提供物体、原始锚点和探索区域的空间信息。通过已观测区域中的持续读出，再接入 Nav2 导航与 Contact-GraspNet 抓取候选，将重建、搜索与实机执行连接起来。

Input roles and the empty-anchor bottleneck输入分工与空置锚点搜索

The map supplies observed geometry, while the photograph supplies appearance and context across time. In the same held-out evaluation, shuffling goal photographs, zeroing the online triplane input and removing anchor affinity reduce ordered localization from 34.8% to 1.8%, 0.8% and 8.4%, respectively. These controls show dependence on both the evolving spatial representation and the remembered scene context in this evaluation.地图提供已观测的空间结构，照片提供跨时间的外观与上下文。在同一组未见场景评测中，打乱目标照片、将在线三平面输入置零、移除锚点关联后，顺序定位成功率由 34.8% 分别降至 1.8%、0.8% 和 8.4%。这些对照说明，持续更新的空间表示与目标场景中的上下文都参与了定位。

Across ten real trials, eight reach object localization and five reach anchor localization; those five complete restoration. The largest stage loss is therefore finding the original location. This directs improvement toward contextual matching of the empty support region.十次实机实验中，八次成功定位物体，五次成功定位锚点，并完成复原。阶段间最大的损失发生在原位置搜索。进一步改进的重点因而指向空置支撑区域的上下文匹配。

Docking for the manipulation that follows面向后续操作的停靠选择

Copy-Paste studies precise restoration to a visual reference with a mobile manipulator. Base placement changes the camera view and the arm’s starting state; errors during grasping, transfer and release can propagate to the final object position. A reachable base pose must therefore be judged by the restoration it enables.Copy-Paste 研究移动操作机器人如何依据视觉参考完成精确复原。底盘停靠位置会改变相机视点和机械臂起始状态，抓取、搬运与释放中的误差也可能传递到终态。因此，停靠位置的选择需要考虑它能否支持后续复原，而不只判断机械臂是否可达。

Current point clouds produce docking candidates. A utility predictor trained on complete restoration outcomes of a frozen WAM helps rank feasible poses together with prior scores. Removing utility scoring while retaining the same twelve candidates reduces mean simulation success across three training seeds, each evaluated on the same 150 cases, from 71.1% to 47.3%. This makes manipulation outcomes the supervision for choosing where the robot should begin.方法根据当前点云生成停靠候选，效用预测器以冻结 WAM 的完整恢复结果训练，并与候选先验分数共同排序。在保留相同十二个候选、每个训练种子评估相同 150 个案例的条件下，移除效用评分使三个种子的仿真平均成功率由 71.1% 降至 47.3%。这一设计把后续操作结果作为停靠选择的依据。

Prediction refinement and physical feedback预测修正与真实执行反馈

Before action generation, two recursive rounds refine the predicted future. The robot then executes the first 24 actions, observes again and replans. Prediction refinement addresses the model’s proposed future; new observations address what actually happened during execution. In the simulation inference study, a third round adds little success while increasing planning latency, making the amount of refinement a practical system trade-off.动作生成前，模型进行两轮递归未来预测细化；随后执行前 24 步动作，重新观测并规划。预测细化修正模型预想的未来，新观测则反映真实执行后的状态，两者在不同阶段发挥作用。在仿真推理对照中，第三轮细化的成功率增益已经很小，而规划时延上升，细化深度因而需要与执行系统的响应时间共同考虑。

Multi-UAV planning and mission interfaces多机规划与任务接口

Ten-UAV indoor traversal couples local collision avoidance with the gaps between vehicles. I led EGO-Planner adaptation for planning, avoidance and replanning, adjusting safety margins, velocity constraints and takeoff order for narrow openings, corners and intersecting paths. ROS2 topics and services connect mission initialization, odometry, cameras, trajectories and completion results.十机穿屋需要同时处理局部避障与队伍间隔。我负责 EGO-Planner 的规划、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、速度约束与起飞顺序；通过 ROS2 Topic／Service 接通任务初始化、里程计、相机、轨迹和完成结果。

For three real UAVs, I contributed to FAST-LIO scan-to-map matching, initial alignment and odometry fault investigation, including contributions to increasing LIO updates from 10 Hz to 100 Hz. Hardware integration covered part of the selection, mounting, power wiring, interface configuration and troubleshooting for Livox Mid-360, cameras, Jetson Orin NX, flight controllers and airframes.三机实飞阶段，我参与 FAST-LIO 的 scan-to-map 匹配、初始对齐与里程计异常排查，以及将 LIO 更新频率由 10 Hz 提升至 100 Hz 的联调。硬件工作涉及 Livox Mid-360、相机、Jetson Orin NX、飞控与机架的部分选型、安装固定、供电接线、接口配置和故障排查。

Sensor mounting constrains the planner传感器安装与规划约束

LiDAR mounting changes both localization features and what the planner can observe. The VR1 development material records that forward tilt captures more ground features but introduces a blind zone; matching the real-platform tilt requires planner retuning. Mounting geometry and avoidance parameters therefore need to be considered together.雷达安装方式会同时改变定位可用的特征和规划可见的障碍物。VR1 开发材料记录了前倾雷达带来的地面特征收益与观测盲区，以及保持实机倾角后重新调整规划参数的需求。这使硬件安装几何与避障参数成为需要共同考虑的设计因素。

In VR1, simulated localization noise does not reproduce accumulated drift, and simulated LiDAR does not observe moving drones. Real deployment therefore calls for separate checks of coordinate alignment, drift and returns from other vehicles. These differences explain why completing a simulated route alone is insufficient preparation for reliable indoor flight.VR1 仿真中的随机定位噪声不会复现累积漂移，仿真雷达也不观测移动无人机。迁移到实机时，需要另外检查坐标对齐、漂移及其他无人机的点云回波。这些差异说明，仿真路线验证之外，定位与真实传感器观测仍需要针对性联调。