Research

We advance state-of-the-art spatial intelligence systems.

GSMem: 3D Gaussian Splatting as Persistent Spatial Memory for Zero-Shot Embodied Exploration and Reasoning

We utilize 3DGS serves as a persistent spatial memory for embodied navigation, enabling the agent to ''hallucinate'' optimal views for high-fidelity Vision-Language Model (VLM) reasoning.

Embodied AI 3D Vision VLM Reasoning

3DGS Spatial Memory Zero-Shot Exploration Embodied Reasoning

Reconstruction Matters: Learning Geometry-Aligned BEV Representation through 3D Gaussian Splatting

we propose Splat2BEV, a Gaussian Splatting-assisted BEV perception framework that aims to learn BEV feature representations that are both semantically rich and geometrically precise.

Embodied AI 3D Vision

Autonomous Driving 3DGS BEV Perception

Segment then Splat: Unified 3D Open-Vocabulary Segmentation via Gaussian Splatting

[NeurIPS'25] Segment then Splat: Unified 3D Open-Vocabulary Segmentation via Gaussian Splatting

We propose Segment then Splat, an Open-vocabulary 3D segmentation method that reverses the long established approach of “segmentation after reconstruction” by dividing Gaussians into distinct object sets before reconstruction.

3D Scene Understanding 3DGS Open-Vocabulary Segmentation

Fix False Transparency by Noise Guided Splatting

[NeurIPS'25] Fix False Transparency by Noise Guided Splatting

We propose Noise Guided Splatting, a method that handles the inherit “false transparency” artifact in 3DGS by injecting opaque noise Gaussians in the object volume during training, the object surfaces are encourages surface Gaussians to adopt higher opacity.

3D Reconstruction 3DGS

Spatial Intelligence in Vision-Language Models: A Comprehensive Survey

A comprehensive survey addressing how VLMs currently lack spatial intelligence, covering recent advances, taxonomies, and evaluations toward building spatially intelligent AI.

VLM Reasoning Survey

Spatial Intelligence

Balancing Fidelity and Diversity: Synthetic Data Could Stand on the Shoulder of the Real in Visual Recognition

Balancing Fidelity and Diversity: Synthetic Data Could Stand on the Shoulder of the Real in Visual Recognition

Investigates how data fidelity and diversity affect recognition performance through synthetic data curation, offering training-free improvements for visual recognition tasks.

Generative Models

Synthetic Data Data Curation

BARD-GS: Blur-Aware Reconstruction of Dynamic Scenes via Gaussian Splatting

[CVPR'25] BARD-GS: Blur-Aware Reconstruction of Dynamic Scenes via Gaussian Splatting

BARD-GS is a novel approach for robust dynamic scene reconstruction that effectively handles blurry inputs and imprecise camera poses.

3DGS Dynamic Scene Reconstruction Blur Robustness

CAUSAL3D: A Comprehensive Benchmark for Causal Learning from Visual Data

CAUSAL3D: A Comprehensive Benchmark for Causal Learning from Visual Data

A 19-dataset benchmark evaluating causal reasoning capabilities in vision models, revealing that performance drops sharply as causal complexity increases.

VLM Reasoning 3D Vision Benchmarking

Visual Causality

NEBULA: A Unified Framework for Evaluating Embodied AI Systems

A unified ecosystem for evaluating embodied AI systems beyond coarse task success metrics, combining capability tests for fine-grained skill diagnosis and stress tests for robustness under real-world perturbations.

Embodied AI Benchmarking

RT-LTP: Real-Time Latent Trajectory Prediction via Efficient Online Adaptation

RT-LTP: Real-Time Latent Trajectory Prediction via Efficient Online Adaptation

A real-time trajectory prediction framework that reformulates forecasting as a latent-space alignment problem. Achieves up to 54% faster adaptation and 9.9% higher accuracy across multiple autonomous driving benchmarks.

Autonomous Driving Trajectory Prediction Online Adaptation

View-consistent Object Removal in Radiance Fields

View-consistent Object Removal in Radiance Fields

we introduce a novel RF editing pipeline that significantly enhances consistency by requiring the inpainting of only a single reference image. This image is then propagated across multiple views using a depth-based approach, to maintain consistencies.

3D Editing NeRF Object Removal Multi-view Consistency

No research projects match the current search and filter settings.