VU Lab

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

2026-04-03T00:00:00+00:00

This placeholder project studies how embodied agents can combine vision, language, and action context to build richer scene representations in unstructured environments.

Current directions include long-tail object understanding, semantic grounding under ambiguity, and robust multimodal fusion for agents that must act with incomplete observations.

This page is a placeholder for future project details, papers, demos, and datasets.

Spatial Memory for Long-Horizon Embodied Agents

2026-04-02T00:00:00+00:00

This placeholder highlight captures a research direction around persistent world memory for embodied systems.

The goal is to let an agent remember what it has seen, retrieve relevant experiences, and use those memories to guide future actions.

This page can later be replaced with project updates, results, videos, or paper links.

Visual Understanding Benchmark for Open-World Scenes

2026-04-01T00:00:00+00:00

This placeholder highlight summarizes a lab effort focused on benchmarking visual understanding systems under realistic scene complexity.

The project studies how perception models behave when scenes contain clutter, rare objects, ambiguous language, and shifting context.

We use this page as a placeholder for future highlight content on datasets, evaluation protocols, and model analysis.

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

2026-03-30T00:00:00+00:00

This placeholder project studies how embodied agents can combine vision, language, and action context to build richer scene representations in unstructured environments.

Current directions include long-tail object understanding, semantic grounding under ambiguity, and robust multimodal fusion for agents that must act with incomplete observations.

This page is a placeholder for future project details, papers, demos, and datasets.

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

2026-03-29T00:00:00+00:00

This placeholder project focuses on how embodied agents maintain useful spatial memory over long time horizons while reasoning about goals, constraints, and uncertainty.

We are interested in navigation policies, memory-augmented world models, and planning systems that remain effective when tasks require multi-step reasoning across large spaces.

This page is a placeholder for future project details, papers, demos, and datasets.

Robust 3D Mapping and Adaptive Navigation

2026-03-28T00:00:00+00:00

This placeholder project explores how agents can build stable 3D maps while environments evolve, sensing degrades, or scene geometry changes over time.

Representative directions include geometry-aware learning, map refinement under uncertainty, and adaptive navigation policies that remain effective in dynamic real-world deployments.

This page is a placeholder for future project details, papers, demos, and datasets.

Video-Language Grounding for Open-World Agents

2026-03-27T00:00:00+00:00

This placeholder project examines how agents align visual observations with language over time, especially when scenes, objects, and goals evolve beyond closed-set assumptions.

We are interested in open-world recognition, grounded language understanding, and long-horizon video reasoning for systems that operate continuously rather than on isolated clips.

This page is a placeholder for future project details, papers, demos, and datasets.

Foundation Models for Robotic Manipulation

2026-03-26T00:00:00+00:00

This placeholder project explores how large-scale models can support manipulation through reusable skills, structured task abstractions, and more transferable control interfaces.

We are especially interested in how vision-language priors and action representations can improve generalization across tasks, objects, and environments.

This page is a placeholder for future project details, papers, demos, and datasets.

Uncertainty Estimation and Robust Deployment

2026-03-25T00:00:00+00:00

This placeholder project investigates how autonomous systems should estimate confidence, react to uncertainty, and remain reliable when real-world conditions depart from training assumptions.

Representative themes include distribution shift detection, calibrated decision making, and risk-aware inference pipelines for perception and control.

This page is a placeholder for future project details, papers, demos, and datasets.

Interactive 3D World Models for Embodied Training

2026-03-24T00:00:00+00:00

This placeholder project focuses on compact world representations that can support simulation, future prediction, and scalable training of embodied agents.

We are interested in interactive 3D environments, controllable simulators, and learned world models that bridge synthetic training and real deployment.

This page is a placeholder for future project details, papers, demos, and datasets.