Move to Understand a 3D Scene: Bridging Visual Grounding and Exploration for Efficient and Versatile Embodied Navigation

2025-07-05

AI Summary

Existing 3D vision-language models are limited to static object localization and lack active exploration capabilities. This paper proposes an online active perception framework for embodied agents, unifying visual grounding and exploration decision-making for the first time. Our approach constructs spatial memory via online query representation learning—eliminating explicit 3D reconstruction—and introduces an end-to-end trajectory learning paradigm that jointly processes RGB-D inputs and leverages vision-language-exploration multimodal pretraining. We optimize navigation policies using over one million simulated and real-world trajectory samples. Evaluated on HM3D-OVON, GOAT-Bench, and other benchmarks, our method achieves absolute improvements of 14%, 23%, 9%, and 2% in task success rates, respectively, significantly enhancing generalizable multimodal navigation performance in complex, dynamic environments.

Technology Category

• Natural Language Processing: Language Grounding & Multi-modal NLP
• Computer Vision: Multi-modal Vision
• Intelligent Robots: Multimodal Perception & Sensor Fusion

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Embodied scene understanding requires not only comprehending visual-spatial information that has been observed but also determining where to explore next in the 3D physical world. Existing 3D Vision-Language (3D-VL) models primarily focus on grounding objects in static observations from 3D reconstruction, such as meshes and point clouds, but lack the ability to actively perceive and explore their environment. To address this limitation, we introduce underline{ extbf{M}}ove underline{ extbf{t}}o underline{ extbf{U}}nderstand ( extbf{model}), a unified framework that integrates active perception with underline{ extbf{3D}} vision-language learning, enabling embodied agents to effectively explore and understand their environment. This is achieved by three key innovations: 1) Online query-based representation learning, enabling direct spatial memory construction from RGB-D frames, eliminating the need for explicit 3D reconstruction. 2) A unified objective for grounding and exploring, which represents unexplored locations as frontier queries and jointly optimizes object grounding and frontier selection. 3) End-to-end trajectory learning that combines extbf{V}ision- extbf{L}anguage- extbf{E}xploration pre-training over a million diverse trajectories collected from both simulated and real-world RGB-D sequences. Extensive evaluations across various embodied navigation and question-answering benchmarks show that MTU3D outperforms state-of-the-art reinforcement learning and modular navigation approaches by 14%, 23%, 9%, and 2% in success rate on HM3D-OVON, GOAT-Bench, SG3D, and A-EQA, respectively. model's versatility enables navigation using diverse input modalities, including categories, language descriptions, and reference images. These findings highlight the importance of bridging visual grounding and exploration for embodied intelligence.

Problem

• Integrates active perception with 3D vision-language learning
• Enables embodied agents to explore and understand environments
• Improves success rates in navigation and question-answering benchmarks

Innovation

• Online query-based representation learning from RGB-D frames
• Unified objective for grounding and exploring via frontier queries
• End-to-end Vision-Language-Exploration pre-training on trajectories