Скачать или смотреть [PhD Defense] Chao Chen: Topological and Metric Spatial Representation for Embodied Agent

Embodied AI systems face significant challenges in perceiving complex environments accurately and reasoning about spatial relationships effectively, particularly in dynamic and unstructured settings. These challenges hinder reliable planning, decision-making, and autonomous interaction.

Mapping plays a crucial role in overcoming these challenges by providing different levels of environmental representation. Metric maps are essential for detailed tasks like collision avoidance and fine-grained planning, offering spatial accuracy at a granular level. In contrast, topological maps provide high-level abstractions that are well-suited for global understanding and long-term reasoning.

This dissertation presents a series of studies addressing distinct perceptual and cognitive aspects of environmental mapping for Embodied AI. It begins with self-supervised metric mapping, enhancing precise and detailed spatial perception crucial for fine-grained navigation and interaction tasks. Building upon this, it develops self-supervised topological mapping techniques that capture high-level, abstract environmental structures, enabling rapid spatial awareness and adaptation in large-scale and diverse scenarios. Finally, it explores how topological maps can support advanced spatial reasoning and cognitive inference, allowing agents to understand complex co-visibility relationships across observations and improve decision-making under uncertainty.

Through extensive experiments on multiple datasets and modalities, this work demonstrates the complementary roles of metric and topological maps in providing Embodied AI systems with both microscopic and macroscopic environmental understanding. The proposed methods significantly improve mapping accuracy, autonomous labeling capability, and spatial reasoning robustness, collectively advancing the state of Embodied AI.

These contributions lay a foundation for future exploration of multi-scale mapping and reasoning approaches, guiding the development of more versatile, generalizable, and cognitively inspired autonomous agents.