About
Liu Yi 刘弈
M.Sc. Computer Engineering · National University of Singapore
🔍 Seeking research internship opportunities in robotics & AIAbout Me
I am a Master's student in Computer Engineering at the National University of Singapore (NUS), with a strong interest in robot autonomy and intelligent systems. My work spans multi-sensor navigation (thermal, LiDAR, 4D radar), SLAM, end-to-end learning for robotic control, 3D visual perception, and real-world system deployment on UGV and humanoid robot platforms.
I enjoy bridging learning-based methods with practical robotic systems — getting things to actually run on hardware is what excites me most.
Research Interests
- Multi-sensor robot navigation — thermal, LiDAR, 4D mmWave radar
- SLAM and autonomous exploration in unknown environments
- 3D visual perception and scene reconstruction
- End-to-end learning for robotic perception and control
- Embodied AI and humanoid robot systems
Education
National University of Singapore
M.Sc. Computer Engineering · GPA 4.7 / 5.0
Robotics and Embodied AI (A+) · Machine Learning with Applications (A)
Aug 2025 – Dec 2026
Singapore
Singapore
Beijing Jiaotong University
B.Eng. Computer Science · GPA 3.76 / 4.0
First-Class Scholarship · Outstanding Graduate Thesis
Sep 2021 – Jun 2025
Beijing, China
Beijing, China
Skills
Languages
Tools & Env
Sensors
Methods
Experience
Multi-Sensor Robot Navigation Research Research Aug 2025 – Dec 2025
NUS CORE Lab
- Built a stereo thermal / RGB-D / LiDAR data acquisition pipeline with hardware-level time synchronization and calibration for low-light and visually degraded environments.
- Integrated BridgeDepth, NVBlox, Nav2, and SLAM pose input on NVIDIA AGX Orin for point-goal navigation on a UGV; validated the full system on real hardware.
- Connected RPM-Net (scan-to-scan soft correspondences) and Radar-Transformer (scan-to-submap hard correspondences) into a radar-inertial odometry backend (IEKF / VGICP); evaluated on ColoRadar and SNAIL datasets.
DRL-based Indoor Autonomous Exploration Research Dec 2025 – Present
NUS CORE Lab
- Identified failure modes of TARE planner in narrow corridors and junctions (goal=NaN, frontier utility collapse); restructured the 3D SLAM → 2D OccupancyGrid perception-planning interface.
- Replaced TARE with ARiADNE (DRL-based explorer) using a minimal-substitution strategy, integrating Livox/Hesai SLAM, OctoMap/OccupancyGrid, and waypoint interfaces while preserving the downstream local planner.
- Built smoke-test and experiment orchestration in Unity and Gazebo; pre-trained checkpoint completed exploration in all 5 test episodes, confirming weight compatibility between training and ROS2 inference repos.
Humanoid Robot Perception & Control Integration Internship Feb 2026 – Mar 2026
Fatfish AI
- Participated in a project building the Fourier N1 visual-following pipeline (RGB-D → remote Orin inference → arm control); studied and documented the algorithm architecture and inference logic.
- Developed the web UI and connected the frontend status display to the backend control system; added RTT monitoring and system diagnostics to support on-site bring-up.
- Used SLAM for environment mapping; resolved camera ownership, network routing, and process scheduling conflicts during live debugging sessions.
3D Perception — Model Fine-tuning & Distillation Internship Apr 2025 – Aug 2025
XiaoYu ZhiZao (Embodied AI startup)
- Fine-tuned MAST3R with low-resolution ROPE and relative position encoding to accelerate inference while maintaining point cloud quality for robot arm planning.
- Designed a distillation loss and applied it to VGGT to speed up image-to-point-cloud inference.
ViT-based Quadrotor End-to-End Obstacle Avoidance Thesis Sep 2024 – May 2025
Bachelor's Thesis · NUS Chongqing Research Institute
- Reproduced ViT / ViT-LSTM baselines from an ICRA 2025 paper; verified superior generalization of ViT in unseen forest environments.
- Rebuilt the expert data pipeline in Flightmare, modified input from single-frame depth to RGB image sequences with an IMU fusion channel.
- Achieved 73.26% obstacle avoidance success rate in forest simulation (Flightmare + Dodgelib).
Publications
ISPRS 2026✓ Extended Abstract — Accepted
Deep Radar Point Matching for 4D Radar-Inertial Odometry
Integrating learned scan-to-scan and scan-to-submap radar correspondences (RPM-Net, Radar-Transformer) into an IEKF/VGICP radar-inertial odometry backend. Evaluated on ColoRadar and SNAIL datasets.
Integrating learned scan-to-scan and scan-to-submap radar correspondences (RPM-Net, Radar-Transformer) into an IEKF/VGICP radar-inertial odometry backend. Evaluated on ColoRadar and SNAIL datasets.
关于我
我目前是新加坡国立大学（NUS）计算机工程专业的硕士生，专注于机器人自主能力与智能系统研究。研究内容涵盖多传感器导航（热红外、LiDAR、4D 雷达）、SLAM、端到端机器人控制学习、3D 视觉感知，以及在 UGV 和人形机器人平台上的实机系统部署。
我的兴趣在于将学习方法与真实机器人系统相结合，让系统能在真实硬件上稳定运行是我最有动力做的事。
研究兴趣
- 多传感器机器人导航——热红外、LiDAR、4D 毫米波雷达
- SLAM 与未知环境自主探索
- 3D 视觉感知与场景重建
- 机器人感知与控制的端到端学习
- 具身 AI 与人形机器人系统
教育背景
新加坡国立大学（QS 第 8）
硕士 · 计算机工程 · GPA 4.7 / 5.0
Robotics and Embodied AI (A+) · Machine Learning with Applications (A)
2025.08 – 2026.12
新加坡
新加坡
北京交通大学（211）
本科 · 计算机科学与技术 · GPA 3.76 / 4.0
一等学习奖学金 · 校级优秀毕业设计
2021.09 – 2025.06
北京
北京
专业技能
编程语言
工具与环境
传感器
算法方向
经历
多传感器机器人导航与定位研究 科研 2025.08 – 2025.12
新加坡国立大学 CORE Lab
- 面向低照度与视觉退化场景，搭建双目热红外 / RGB-D / LiDAR 多传感器采集链路，完成硬件时间同步、内外参标定。
- 在 NVIDIA AGX Orin 上集成 BridgeDepth、NVBlox、Nav2 与 SLAM 位姿输入，实现 UGV 上的 point-goal 导航并完成实机验证。
- 将 RPM-Net（scan-to-scan 软对应）与 Radar-Transformer（scan-to-submap 硬对应）接入雷达惯性里程计后端（IEKF / VGICP），在 ColoRadar 与 SNAIL 数据集上完成评测。
基于 DRL 的室内自主探索导航系统重构 科研 2025.12 – 至今
新加坡国立大学 CORE Lab
- 分析 TARE 规划器在窄走廊、路口场景下的 goal=NaN 与 frontier utility 失效问题，梳理 3D 点云 SLAM 到 2D 占用栅格的感知-规划接口。
- 采用"最小替换"策略将 ARiADNE（DRL 自主探索器）接入现有 ROS2 导航链路，打通 Livox/Hesai SLAM、OctoMap/OccupancyGrid 与 waypoint 接口，保留下游 local planner。
- 在 Unity 与 Gazebo 中建立 smoke test 流程；预训练 checkpoint 在 5 个测试 episode 上均完成探索，验证了训练仓与 ROS2 推理仓的权重兼容性。
人形机器人感知与控制联调 实习 2026.02 – 2026.03
Fatfish AI
- 参与 Fourier N1 视觉跟随系统的开发（RGB-D → Orin 远程推理 → 手臂控制），理解并梳理视觉跟随算法的架构与推理逻辑。
- 负责 Web UI 开发，打通前端状态展示与后端控制系统的连接；增加 RTT 监控与状态诊断模块，用于现场联调。
- 使用 SLAM 进行环境建图；处理相机归属、网络配置与进程调度等现场调试问题。
3D 感知——端到端重建模型微调与蒸馏 实习 2025.04 – 2025.08
小雨智造（具身智能初创公司）
- 对 MAST3R 进行低分辨率 ROPE 微调和相对位置编码微调，在保持点云质量的同时加速推理，用于机械臂规划场景。
- 设计蒸馏 loss，对 VGGT 进行蒸馏，提升图像到点云的推理速度。
基于 ViT 的四旋翼无人机端到端避障系统 毕设 2024.09 – 2025.05
本科毕业设计 · 新加坡国立大学重庆研究院
- 复现 ICRA 2025 论文中的 ViT / ViT-LSTM 基线，验证 ViT 在陌生仿真环境中更优的泛化性。
- 基于 Flightmare 重构专家数据采集流程，将输入从单帧深度图改为 RGB 图像序列并引入 IMU 融合通道。
- 在树林仿真环境（Flightmare + Dodgelib）中实现 73.26% 避障成功率。
论文
ISPRS 2026✓ Extended Abstract — 已接受
Deep Radar Point Matching for 4D Radar-Inertial Odometry
将 RPM-Net（scan-to-scan 软对应）与 Radar-Transformer（scan-to-submap 硬对应）接入 IEKF/VGICP 雷达惯性里程计，在 ColoRadar 与 SNAIL 数据集上完成评测。
将 RPM-Net（scan-to-scan 软对应）与 Radar-Transformer（scan-to-submap 硬对应）接入 IEKF/VGICP 雷达惯性里程计，在 ColoRadar 与 SNAIL 数据集上完成评测。