Hongbo Duan

I'm a second-year Ph.D. student at the Tsinghua University Shenzhen International Graduate School, where I work in the Intelligent Robotics Lab. My research interests lie in 3D computer vision and embodied intelligence, with a focus on 3D Gaussian Splatting, neural scene reconstruction, and autonomous navigation.

Email  /  Scholar  /  Github

Research

I work on 3D vision, embodied intelligence, and generative scene modeling. My recent work explores efficient 3D Gaussian Splatting for dynamic and active reconstruction, single-image 3D scene generation, and long-horizon embodied navigation in open-world scenarios.

An active reconstruction framework based on 3DGS for dynamic scenes. It decouples structural and motion uncertainty to guide view selection and path planning, forming a closed-loop active reconstruction pipeline that outperforms existing baselines in accuracy and exploration efficiency.

A single-image 3D scene exploration framework that combines video diffusion priors with geometry-aware expansion strategies. It features a camera trajectory modulation module and a collision-aware planner for controllable large-viewpoint novel view synthesis, and achieves SOTA 3DGS reconstruction on RealEstate10K.

An open-world semantic navigation framework driven by geodata augmentation and RAG. It builds multi-level embodied graphs and robust planning algorithms to solve long-distance navigation in dynamic outdoor environments, supporting dynamic retrieval of both coarse buildings and fine-grained objects.

GSWAM is a 3D Gaussian‑based world‑action model. It unifies scene dynamics and robot actions in a shared 3D latent space, enabling joint action generation and scene prediction, and achieves strong results on simulated and real‑world robotic manipulation.

G2P-WAM is a geometry-to-preference post-training framework that connects static–dynamic representation distillation to action-policy preference alignment. It features three stages: GeoSFT grounds WAM representations with removable static-3D and dynamic-4D supervision, reward diagnosis validates geometric scores on the policy's own rollouts, and GeoDPO aligns the action stream with same-context preference pairs. It improves closed-loop manipulation success across benchmarks with no inference overhead.

We propose STaR‑WAM, a training‑only geometric supervision framework for World‑Action Models. Static and temporal geometry alignment losses distill 3D structural and motion knowledge without manual annotations. All auxiliary teachers are removed after training, bringing no inference overhead. It achieves higher success rates on geometry‑sensitive robotic manipulation tasks and faster convergence.

A 3D Gaussian‑based visual SLAM framework for dynamic environments. It eliminates the need for predefined semantics or depth inputs, exploits sequential attention and dynamic feature enhancement to separate static‑dynamic components, and adopts occlusion filling and distractor‑adaptive SSIM loss for strong robustness, achieving state‑of‑the‑art localization and dense reconstruction results.

EventGS is a near real‑time 3DGS‑based SLAM system for monocular event streams. It leverages diffusion‑restored intensity and event‑accumulated observations together with progressive refinement to enable accurate pose tracking and high‑quality renderable reconstruction under extreme HDR and fast‑motion conditions.

LIO‑Track is a tightly‑coupled LiDAR‑Inertial odometry with multi‑object tracking built upon hierarchical factor‑graph optimization and ESKF, which jointly optimizes SLAM and MOT for improved pose accuracy and robust dynamic‑object processing.

Gen3d generates high‑quality generic 3D Gaussian‑based world models from a single image. It lifts RGBD input into point clouds, expands and optimizes the scene representation to synthesize consistent, high‑fidelity novel views.

We introduce a Fourier motion modeling module for 4D Gaussian Splatting, decomposing motion into sinusoidal frequency components to capture complex dynamics. Combined with frequency‑aware regularization, it enhances complex‑motion fitting and long‑term temporal coherence while preserving real‑time rendering.

Template based on Jon Barron's website.