In the field of Embodied AI, agents are tasked with navigating complex 3D environments to find objects or follow natural language instructions. Traditionally, these agents have relied on "latent memory"—internal vector states like those in LSTMs—to keep track of where they have been. However, as Chen Gao and colleagues point out in Reinforced Structured State-Evolution for Vision-Language Navigation, compressing an entire 3D layout into a one-dimensional vector often leads to "forgetting." The agent loses the spatial relationships between objects, resulting in redundant exploration or circular paths.
To solve this, a growing body of research has shifted toward Topological Graph Memory. Instead of a single vector, the agent builds an explicit map where nodes represent visited locations and edges represent the navigable paths or geometric relationships between them.
From Sequences to Structures
The core intuition behind topological memory is that spatial reasoning is fundamentally different from sequence processing. In Graph Convolutional Memory using Topological Priors, researchers argue that organizing observations by spatial proximity rather than temporal order provides a much stronger inductive bias for navigation.
A landmark approach in this area is the Semi-Parametric Topological Memory (SPTM) proposed by Nikolay Savinov in Semi-parametric topological memory for navigation. SPTM combines a non-parametric graph of environmental observations with a deep neural network for local control. By establishing "shortcuts" between visually similar nodes that might have been visited at different times, the agent transforms a linear history into a dense, navigable graph. This allows for global planning that is robust to sensor drift and does not require precise metric coordinates.
Navigating the Long Horizon
Standard reinforcement learning models often fail as the distance to a goal increases. Dinesh Jayaraman and his team addressed this in Long-HOT: A Modular Hierarchical Approach for Long-Horizon Object Transport. They demonstrated that a hierarchical policy—using a topological graph to store object locations and vantage points—allows agents to backtrack and plan over long horizons where reactive models typically degrade.
Similarly, Hanqing Wang 0001 et al. introduced Structured Scene Memory for Vision-Language Navigation, which uses message-passing over a topological graph. This mechanism allows the agent to aggregate information from neighboring nodes, giving it a "global perspective" that extends beyond its current field of view. By storing visual features at nodes and geometric relationships (like heading and elevation) at edges, the agent maintains a persistent record of the environment's connectivity.
Semantic and 3D Scene Graphs
Modern navigation tasks often require more than just knowing "where" a path leads; they require understanding "what" is in the room. This has led to the development of 3D Scene Graphs. In SayNav: Grounding Large Language Models for Dynamic Planning to Navigation in New Environments, Abhinav Rajvanshi describes an agent that incrementally builds a semantic graph of objects and rooms. This graph serves as a structured prompt for a Large Language Model (LLM), allowing the agent to use "common sense" to decide where a target object is likely to be found.
This semantic grounding is further explored by Luca Weihs in Scene Graph Contrastive Learning for Embodied Navigation. By training agents to align their internal beliefs with a structured scene graph, the model learns to remember objects even after they are no longer visible, significantly boosting performance in object-seeking tasks.
Persistence Across Episodes
Perhaps the most significant advantage of graph-based memory is its ability to persist across multiple tasks. While traditional agents treat every new instruction as a "blank slate," researchers like Haodong Hong and Qi Wu argue in General Scene Adaptation for Vision-and-Language Navigation that a robot living in a home should leverage its previous explorations.
By maintaining a persistent topological graph across episodes, agents can:
- Avoid Redundant Exploration: Fengda Zhu et al. show in SOON: Scenario Oriented Object Navigation With Graph-Based Exploration that graphs prevent circular movement by tracking visited vs. unvisited nodes.
- Solve the Cold-Start Problem: Yongqiang Yu et al. demonstrate in User-Feedback-Driven Continual Adaptation for Vision-and-Language Navigation that caching visual features and topology allows agents to navigate more efficiently in subsequent interactions within the same environment.
- Reason with Imagination: In Planning from Imagination: Episodic Simulation and Episodic Memory for Vision-and-Language Navigation, Yuanlong Pan and colleagues even integrate "imagined" nodes—predicted layouts of unvisited rooms—into the topological map to help the agent plan before it even enters a space.
By moving from flat, temporal memories to structured, spatial graphs, the field is enabling agents to navigate not just as reactive controllers, but as informed explorers with a persistent sense of place.