Zero-Shot Spatial Reasoning for Embodied Navigation

This direction is split into other directions: zero-shot spatial reasoning. Its last version stays readable and citable here.

Version 1 of 1121 papers148 conceptsfrozen 2026-10-04

Traditional robotic navigation relies on "data-driven" paradigms: an agent is trained in thousands of simulated environments to recognize patterns and map visual inputs to motor commands. However, these models often fail when faced with "long-tail" scenarios or novel environments that differ even slightly from their training data.

A new direction in embodied AI shifts the burden of navigation from task-specific training to the inherent "world knowledge" of Large Language Models (LLMs) and Vision-Language Models (VLMs). By treating navigation as a zero-shot reasoning problem, researchers are enabling robots to explore unfamiliar spaces, find novel objects, and cross complex terrains without any environment-specific fine-tuning.

The Common Sense Prior

The core intuition behind this shift is that foundation models already "know" the world. An LLM understands that a "leopard-print chair" is likely in a living room and that a "milk carton" is likely in a refrigerator. This semantic common sense acts as a powerful heuristic for exploration.

In IPPON: Common Sense Guided Informative Path Planning for Object Goal Navigation, Kaixian Qu and colleagues demonstrate how LLMs can derive spatial relationships between "anchor objects" (like a stove) and a target (like a spatula). These linguistic proximity rankings are projected into a spatial map, biasing the robot's exploration toward areas with high semantic likelihood. Similarly, Jinzhou Lin and Li Guo highlight in Advances in Embodied Navigation Using Large Language Models: A Survey that LLMs serve as a repository of world knowledge, allowing agents to infer object locations in novel settings by applying broad semantic context to specific topological structures.

Hierarchical Reasoning and Control

While LLMs excel at high-level strategy, they cannot directly control a robot's motors. The emerging solution is a hierarchical framework: the LLM acts as the "planner," decomposing complex goals into sub-tasks, while a low-level controller (often trained via reinforcement learning) handles the physical execution.

Shaoting Zhu and the team behind SARO: Space-Aware Robot System for Terrain Crossing via Vision-Language Model use this approach to help quadruped robots cross 3D obstacles like stairs and gaps. The VLM identifies traversable paths—such as a ramp—and decomposes the navigation goal into executable steps. This high-level reasoning is bridged to the real world through a "Probability Annealing Selection" (PAS) method, ensuring stable locomotion across challenging terrains. This same hierarchical logic is seen in L2M2: A Hierarchical Framework Integrating Large Language Model and Multi-agent Reinforcement Learning, where Minghong Geng uses LLMs to steer agents away from environmental traps like cliffs, offloading the "common sense" of safety to the pre-trained model.

Overcoming the Local Minimum with Spatial Memory

A significant challenge for zero-shot agents is "local-only" thinking—the tendency to get stuck in dead ends because they lack a persistent memory of the environment. To solve this, researchers are equipping LLMs with structured, text-based representations of space.

MapGPT: Map-Guided Prompting for Unified Vision-and-Language Navigation by Bingqian Lin introduces an online-constructed topological map that allows the agent to perform multi-step adaptive path planning. By "verbalizing" the environment into a graph of connected nodes, the LLM can reason about where it has already been and where it should go next. Daojie Peng further explores this in TopoNav: Topological Graphs as a Key Enabler for Advanced Object Navigation, showing that dynamic topological graphs allow robots to backtrack efficiently and reason about room connectivity, mimicking human-like navigation patterns.

Physical Grounding and Traversability

Beyond finding objects, zero-shot models are beginning to understand the physical constraints of the world. This is known as "traversability"—the ability to judge whether a specific terrain is safe for a specific robot.

In ZeST: an LLM-based Zero-Shot Traversability Navigation for Unknown Environments, Shreya Gummadi demonstrates that multimodal LLMs can predict terrain safety by reasoning about the relationship between a robot’s physical characteristics and environmental features (e.g., "this slope is too steep for my wheels"). This capability allows for safe deployment in hazardous, unknown environments without the need for dangerous trial-and-error data collection.

From Bird's-Eye to Situated Perspectives

A final frontier in this research is moving from "bird's-eye" map reasoning to "situated" egocentric reasoning. Yue Zhang and colleagues developed SPARTUN3D to train models on egocentric spatial narratives. By teaching models to describe scenes from the perspective of an agent moving through them, they improve the model's ability to distinguish between visually similar paths and execute precise maneuvers. This "situated" understanding is what allows agents like VELMA to navigate real-world city streets by verbalizing landmarks and intersections into descriptive sentences that an LLM can process.

By combining semantic common sense, hierarchical planning, and structured memory, this research direction is transforming foundation models from passive observers into active, embodied navigators capable of operating in a world built for humans.

Papers in this version

Showing 60 of 121 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_zero_shot_spatial_reasoning_for_embodied_c34653e2,
  title = {Zero-Shot Spatial Reasoning for Embodied Navigation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/zero-shot-spatial-reasoning-for-embodied-navigation/rdr_c34653e2e58b44b1a1d4d9796cb4db5d},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_zero_shot_spatial_reasoning_for_embodied_c34653e2_v1,
  title = {Zero-Shot Spatial Reasoning for Embodied Navigation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/zero-shot-spatial-reasoning-for-embodied-navigation/rdr_c34653e2e58b44b1a1d4d9796cb4db5d/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}