Graph Based Spatial Representations for Embodied AI

515 concepts359 papers1 proposal

Traditional deep reinforcement learning agents often suffer from a "myopic" perspective. Whether navigating a 3D maze or a text-based adventure game, these agents typically rely on recurrent neural networks to maintain a sense of history. However, as Ammanabrolu et al. point out, these architectures frequently struggle with "forgetting" in long-horizon tasks. When an agent turns its virtual head or moves to a new room, the visual or textual features of its previous location often fade from its internal state.

To solve this, a growing body of research proposes representing the environment not as a sequence of observations, but as a persistent, structured graph. By mapping the world into nodes (locations or objects) and edges (spatial or semantic relationships), agents can maintain a "mental map" that persists across time, enabling more human-like reasoning and navigation.

The Persistence of Topological Memory

The core intuition behind graph-based representations is that they decouple an agent's memory from its immediate perception. In the ESceme framework, Liu et al. demonstrate how storing pooled visual features in a persistent scene graph allows agents to leverage information gathered during previous tasks in the same environment. This transforms the agent from a "goldfish" that treats every instruction as a brand-new experience into a system that can recognize previously visited rooms.

This structural approach also serves as an efficient information bottleneck. In text-adventure games, Liu et al. show that factorizing the world into a knowledge graph of discrete object nodes filters out irrelevant textual noise while maintaining a memory of entity states across partially observable sequences. Similarly, Ammanabrolu et al. use these graphs to prune massive action spaces, masking out impossible actions by checking the agent's current "reachability" within its internal map.

Connectivity Over Coordinates

One of the most powerful facets of topological mapping is its robustness to sensor noise. Unlike metric maps, which rely on precise coordinates and are susceptible to drift, topological maps focus on reachability. The Semi-Parametric Topological Memory (SPTM) architecture, described in a paper from ICLR 2018, combines a non-parametric graph of observations with deep retrieval networks. By focusing on whether one landmark is reachable from another rather than their exact distance, the authors achieved near-perfect success rates in unseen 3D environments.

This shift from geometry to connectivity is echoed in cognitive robotics. Smart argues that spatial memory is effectively structured as a directed graph where nodes are defined by unique combinations of sensory inputs. In this view, navigation is not a matter of calculating vectors, but of propagating activation through a network of associated memory nodes to find a path to a target state.

Grounding Reasoning in Scene Graphs

The rise of Large Language Models (LLMs) has provided agents with high-level "common sense," but these models lack physical grounding. Researchers have bridged this gap by using 3D scene graphs as a translation layer. Rajvanshi et al. introduce SayNav, which incrementally builds a 3D scene graph to prompt an LLM for dynamic planning. This allows the agent to update its strategy immediately upon identifying a relevant landmark, such as a specific piece of furniture.

These graphs can also handle hierarchical complexity. For tasks requiring an agent to find "the apple on the table in the kitchen," Lingelbach proposes using scene graphs to filter out environmental noise, facilitating hierarchical reasoning in large-scale homes. By grounding high-level reasoning in these dynamic structures, systems like MoMa-LLM achieve significantly higher search efficiency and reduced travel distances compared to unstructured baselines, as shown by Honerkamp et al..

Global Planning and Error Correction

Perhaps the most significant advantage of a graph-based representation is the ability to perform global reasoning. Myopic agents often get stuck in local loops or fail to recover from navigation errors. The Evolving Graphical Planner (EGP), proposed by Deng et al., maintains a dynamic graph that allows the agent to perform "non-local" transitions. If the current visual input does not match the instructions, the agent can "jump" back to a previously visited node to correct its course.

This global perspective is further enhanced by top-down representations. Liu et al. demonstrate that mapping potential paths to a bird's-eye-view topological map allows agents to distinguish between visually similar paths that lead to different physical locations. By unrolling the environment into a higher-dimensional graph that accounts for topological history, as suggested by Bhattacharya, agents can navigate complex obstacles with a level of foresight that reactive models cannot replicate.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 343 papers
  2. Version 1 · 2026-10-04 · 389 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_topological_environment_representation_dbc55075,
  title = {topological environment representation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/topological-environment-representation/rdr_dbc5507541194011bb052e5af4841277},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 2, frozen text):

@misc{lacuna_topological_environment_representation_dbc55075_v2,
  title = {topological environment representation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/topological-environment-representation/rdr_dbc5507541194011bb052e5af4841277/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake