Wenxuan Guo | 郭文轩

I am a fourth-year Ph.D. student in the Department of Automation at Tsinghua University, advised by Prof. Jianjiang Feng and Prof. Jiwen Lu. In 2022, I obtained my B.Eng. in the XUTELI School, Beijing Institute of Technology.

My current research focuses on 3D vision and embodied intelligence.

Email  /  CV  /  Google Scholar  /  Github

We propose GesVLA, a gesture-aware VLA that embeds hand cues to disambiguate targets in cluttered manipulation scenes. It improves human-robot interaction efficiency and real-robot performance on multiple practical tasks.

We propose AwareVLN, a framework with sparse self-aware reasoning at key navigation nodes via a unified reason-act VLM. An automatic data engine scales structured supervision without manual annotation. It outperforms prior methods on R2R-CE/RxR-CE and deploys on a real quadruped.

We propose IGL-Nav, an incremental 3D Gaussian localization framework for image-goal navigation. It supports challenging scenarios where the camera for goal capturing and the agent's camera have very different intrinsics and poses, e.g., a cellphone and a RGB-D camera.

We propose TSP3D, an efficient multi-level convolution architecture for 3D visual grounding. TSP3D achieves superior performance compared to previous approaches in both accuracy and inference speed.

We propose CL-Gait, the first cross-modality gait recognition framework between camera and LiDAR. We propose a contrastive pre-training method to align the feature spaces of the two modalities, along with a large-scale data generation strategy.

We propose ReID3D, a LiDAR-based ReID framework that utilizes a pre-training strategy to retrieve features of 3D body shape. Additionally, we build LReID — the first LiDAR-based person ReID dataset, which is collected in several outdoor scenes with natural variations.

Selected Projects

GesVLA

Gesture-Aware Robot Manipulation

GesVLA introduces gesture as a parallel instruction modality to disambiguate targets in cluttered scenes and improve human-robot interaction efficiency. Hand cues are encoded in a dual-VLM for joint intent reasoning and action generation, with semi-synthetic data and two-stage training for strong real-robot performance on block, jelly, and produce selection.

AwareVLN

Self-aware Vision-Language Navigation

In AwareVLN, we equip VLN with sparse self-aware reasoning at key nodes via a unified reason–act VLM, together with an automatic data engine for scalable structured supervision. AwareVLN achieves strong results on R2R-CE and RxR-CE with monocular RGB and deploys on a real quadruped.

IGL-Nav

3D Representation for Visual Navigation

We propose incremental 3D Gaussian localization for free-view image-goal navigation in IGL-Nav. We support a challenging application scenario where the camera for goal capturing and the agent's camera have very different intrinsics and poses, e.g., a cellphone and a RGB-D camera.

Website Template