Zhehan Li
Ph.D. Student · FASTLAB (Fire Group), Zhejiang University博士研究生 · 浙江大学 FASTLAB（Fire Group）
College of Control Science and Engineering
Zhejiang University
Hangzhou, China
浙江大学
控制科学与工程学院
中国杭州
I am a Ph.D. student in Control Science and Engineering at Zhejiang University, advised by Yanjun Cao in FASTLAB (Fire Group). I received my B.Eng. in Mechatronics and Automation from Zhejiang University’s Chu Kochen Honors College in 2022 and expect to complete my Ph.D. in June 2027.
My research focuses on multi-robot relative localization and state estimation: enabling robots to determine their poses with respect to one another, then using those estimates for cooperative perception, planning and control. I combine multimodal sensing, geometric estimation and learning-based methods, with an emphasis on algorithms that remain reliable under asynchronous measurements, observation outliers, limited communication and GPS denial. I also build the sensing and robotic systems needed to validate these methods beyond simulation.
Research Interests
- Multi-robot relative localization and state estimation. Bearing–range–inertial fusion, continuous-time and outlier-robust estimation, and zero-shot sim-to-real learning.
- Cooperative perception and active sensing. Air–ground collaboration, active 3D reconstruction, viewpoint planning and online camera–LiDAR calibration.
- Real-world multi-robot systems. Custom sensing hardware, onboard estimation and control, system integration and multi-robot deployment.
Selected projects are grouped below by my role in each one.
For research discussions or potential collaborations, please feel free to contact me at zhehanli66@gmail.com.
我是浙江大学控制科学与工程学院博士研究生，在 FASTLAB（Fire Group）跟随 曹燕军 老师从事研究。2022 年本科毕业于浙江大学竺可桢学院机械电子工程专业，预计于 2027 年 6 月完成博士学业。
我的研究聚焦于多机器人相对定位与状态估计：让机器人准确获知彼此的相对位姿，并将其用于协同感知、规划与控制。我结合多模态感知、几何估计与学习方法，重点研究异步测量、观测外点、通信受限和 GPS 拒止条件下的可靠估计；同时搭建配套感知硬件与机器人系统，在真机上验证算法。
研究方向
- 多机器人相对定位与状态估计。 方位—测距—惯性融合、连续时间与抗外点估计、零样本 Sim-to-Real 学习。
- 协同感知与主动感知。 空地协同、主动三维重建、视角规划与在线相机—激光标定。
- 多机器人真机系统。 自研感知硬件、机载估计与控制、系统集成及多机器人真机部署。
以下项目按照我在其中承担的角色分类展示。
如希望交流研究问题或探讨合作，欢迎邮件联系 zhehanli66@gmail.com。
Research研究
Projects I led as first or co-first author.我以第一作者或共同第一作者主导的研究。
NeuRIO
NeuRIO: Physics-Guided Edge-State Graph Learning for Zero-Shot Multi-Robot Relative Inertial OdometryNeuRIO：面向多机器人零样本相对惯性里程计的物理引导边状态图学习
A causal graph estimator for streaming 6-DoF multi-robot relative poses from bearing, UWB range and inertial measurements. Trained only in simulation, the frozen model transfers directly to 24 real sequences, averaging 0.121 m in position and 2.53° in rotation.面向方位、UWB 测距与惯性观测的因果图估计器，可流式输出多机器人的 6-DoF 相对位姿。模型仅在仿真中训练，冻结参数后直接迁移到 24 段真机序列，平均位置误差 0.121 m、旋转误差 2.53°。
AIOG
To See All: LiDAR-Guided Active Reconstruction via Online Calibration and Gimbal PlanningTo See All：基于在线标定与云台规划的激光引导主动三维重建
A LiDAR-guided active reconstruction framework that treats appearance coverage as an explicit map state. It combines online residual camera–LiDAR calibration with motion-budgeted gimbal view planning and smooth trajectory refinement; the accompanying COOKIE dataset contains 50,006 hardware-synchronized image–scan pairs from 32 scenes.以激光几何为先验、将外观覆盖显式建模为地图状态的主动三维重建框架。系统结合在线相机—激光残差标定、受运动预算约束的云台视角规划和平滑轨迹优化；配套 COOKIE 数据集包含 32 个场景、50,006 对硬件同步图像—扫描数据。
CREPES-X
CREPES-X: Hierarchical Bearing-Distance-Inertial Direct Cooperative Relative Pose Estimation SystemCREPES-X：多层次方位—测距—惯性融合的多机器人直接相对位姿估计系统
A compact infrared-fisheye / UWB / IMU module paired with a hierarchical direct estimator: closed-form single-frame initialization followed by tightly coupled multi-frame refinement. CREPES-X reaches 0.073 m / 1.8° RMSE in real-world tests and maintains real-time operation with 20 robots.将紧凑型红外鱼眼 / UWB / IMU 模块与多层次直接式估计器结合：先以单帧闭式方法快速初始化，再通过紧耦合多帧优化提高精度与鲁棒性。真机实验达到 0.073 m / 1.8° RMSE，并支持 20 台机器人实时运行。
CT-RIO
Parallel Continuous-Time Relative Localization with Augmented Clamped Non-Uniform B-Splines基于增广钳制非均匀 B 样条的并行连续时间相对定位
A continuous-time relative localization framework that fuses asynchronous bearing, range and IMU measurements at their native timestamps while jointly estimating inter-robot clock offsets. Its parallel solver achieves 0.046 m / 1.8° RMSE and reduces error by up to 60% over discrete-time optimization on high-speed sequences.连续时间相对定位框架，可按原始时间戳融合异步的方位、测距与 IMU 观测，并联合估计机间时钟偏移。并行求解器在真机实验中达到 0.046 m / 1.8° RMSE，高速序列上的误差较离散时间优化最高降低 60%。
ColAG
ColAG: A Collaborative Air-Ground Framework for Perception-Limited UGVs’ NavigationColAG：面向感知受限无人车的空地协同导航框架
A single perception-equipped UAV serves as a mobile sensor for UGVs that carry only wheel odometry and relative-localization hardware. Each UGV predicts collision risks from localization uncertainty and unknown space, while the UAV schedules support through a dynamic vehicle-routing formulation; experiments scale to seven UGVs in simulation and three on hardware.由一架具备完整感知能力的无人机为仅搭载轮式里程计和相对定位硬件的无人车提供移动感知。无人车根据定位不确定性与未知区域预测碰撞风险，无人机通过动态车辆路径问题调度支援；系统在仿真中验证至 7 台无人车，并完成 3 台无人车的真机实验。
Co-authored Research合作研究
Projects I contributed to as a co-author.我以合作作者身份参与的研究。
CoNi-PPO
Observation-Based Reinforcement Learning for Preview-Guided UAV Tracking in Non-Inertial Relative Frames基于观测的强化学习：非惯性相对坐标系下的预观测引导无人机跟踪
An end-to-end PPO controller for tracking a moving reference robot in its non-inertial frame. By inferring the reference robot’s IMU state from relative observations, it removes the need for a communication link and reduces high-speed tracking error by 50% relative to CoNi-MPC.面向非惯性相对坐标系跟踪的端到端 PPO 控制器。策略从相对观测中推断参考机器人的 IMU 状态，无需机间通信；在高速场景下，跟踪误差较 CoNi-MPC 降低 50%。
Mr. Virgil
Mr. Virgil: Learning Multi-robot Visual-range Relative LocalizationMr. Virgil：学习式多机器人视觉—测距相对定位
A learning-based solution to the ambiguous association between visual detections and UWB ranges in multi-robot localization. A graph network predicts associations, initial relative positions and their uncertainties; a differentiable pose-graph back end refines them jointly and propagates supervision through the full pipeline.面向多机器人视觉检测与 UWB 测距歧义关联的学习式相对定位方法。图网络预测数据关联、相对位置初值及其不确定度，再由可微位姿图后端联合优化，并将监督信号贯穿整个流程。
Cost-Effective Swarm Navigation
Cost-Effective Swarm Navigation System via Close Cooperation紧密协同的低成本集群导航系统
A tightly coupled air–ground swarm in which one UAV-mounted depth camera provides environmental perception for the entire team. The UAV explores and assists localization while the UGVs optimize collision-adaptive trajectories, raising success in concave-obstacle scenes from 30% to 90% under a limited camera field of view.由一架无人机搭载的深度相机为整个空地集群提供环境感知。无人机负责探索与辅助定位，无人车优化碰撞自适应轨迹；在相机视场受限时，凹形障碍场景的导航成功率由 30% 提升至 90%。
CoNi-MPC
CoNi-MPC: Cooperative Non-Inertial Frame Based Model Predictive ControlCoNi-MPC：基于非惯性坐标系的协同模型预测控制
A nonlinear model predictive controller formulated directly in a moving target’s non-inertial frame, without global state estimates or assumptions about the target’s motion. The same controller supports leader following, orbiting, moving-platform landing and prescribed relative trajectories.直接在运动目标的非惯性坐标系中建立四旋翼非线性模型预测控制，无需全局状态，也不预设目标的运动模式。同一控制器可完成领航跟随、环绕飞行、移动平台降落和指定相对轨迹跟踪。
CREPES
CREPES: Cooperative RElative Pose Estimation SystemCREPES：协作式相对位姿估计系统
A cooperative relative-pose sensing system built from coded active infrared LEDs, an infrared fisheye camera, UWB and an IMU. Optical identity coding associates visual bearings with range measurements, allowing 6-DoF relative poses to be recovered from a single measurement frame. Across the evaluated sequences, median position and orientation errors range from 0.102–0.161 m and 0.733–1.517°.由编码式主动红外 LED、红外鱼眼相机、UWB 与 IMU 构成的协作相对位姿感知系统。红外光编码为方位观测提供身份信息，从而完成视觉与测距的数据关联，并由单帧观测恢复 6-DoF 相对位姿。各测试序列的位置与姿态中位误差分别为 0.102–0.161 m 和 0.733–1.517°。
Engineering工程实践
Systems I led or helped build and deploy on real robots.我主导或参与搭建并完成真机部署的系统。
IROS 2025 EXPO DemoIROS 2025 EXPO 现场演示
Air-Ground Collaboration without Global Information无全局信息下的空地协同
I led the integration, hardware testing and on-site deployment of a GPS- and map-free air–ground collaboration system built on CREPES relative localization and CoNi-MPC control. During the public demo, visitors teleoperated a tracked vehicle while the UAV followed it autonomously using only inter-robot observations.我负责无 GPS、无全局地图空地协同系统的平台集成、真机联调与现场部署，系统基于 CREPES 相对定位和 CoNi-MPC 控制。公开演示中，观众实时遥控履带车，无人机仅依靠机间相对观测自主跟随。
Micro-UAV Substation Inspection微型无人机配电房巡检
Autonomous Inspection of Power Distribution Rooms with a Micro UAV微型无人机的配电房自主巡检
A micro-UAV inspection system that uses a single RGB-D camera and runs ORB-SLAM3, EGO-Planner and PX4 on a Jetson Orin NX. It connects offline mapping, waypoint teaching, cross-session relocalization, obstacle avoidance and directed image capture, with safeguards for pose delay and SLAM failure. I lead requirements, system integration and delivery, and initiated one patent application.仅用一颗 RGB-D 相机的微型无人机自主巡检系统，在 Jetson Orin NX 上集成 ORB-SLAM3、EGO-Planner 与 PX4，贯通离线建图、航点示教、跨会话重定位、自主避障和定向拍照，并加入位姿延迟补偿与 SLAM 跟丢保护。我负责需求拆解、系统集成、进度推进与交付，并主导一项专利申请。