How Foundation Models Are Transforming Embodied Navigation from Task‑Specific to General Intelligence

1. Problem Definition and Taxonomy

Embodied navigation is defined as an agent operating in partially observable environments, perceiving its surroundings, understanding navigation goals, making sequential decisions, and executing physical actions to reach target locations. The survey classifies existing work by task type (semantic, geometric, interactive, composite/general navigation) and robot embodiment (wheeled, legged, aerial), analyzing differences in sensing setups, motion constraints, and planning complexity.

2. System Design Driven by Foundation Models

The authors decompose the navigation pipeline into three core modules:

Observation & Representation : basic egocentric RGB, depth, multi‑view inputs; map‑enhanced spatial representations; camera intrinsics/extrinsics for cross‑view alignment.

Memory Mechanisms : visual memory (historical visual context), textual memory (language summaries), and map‑enhanced memory (explicit spatial structures).

Decision & Control : semantic goal selection, discrete action prediction, continuous action generation; decision mechanisms include explicit reasoning, adaptive reasoning, and training‑only supervision.

Four representative architecture paradigms are compared: modular systems with explicit perception‑mapping‑planning‑control separation, single‑policy systems mapping multimodal input directly to actions, dual‑system designs separating high‑level slow reasoning from low‑level fast control, and world‑model‑driven systems that predict future states or environment changes to enhance planning.

3. Data Collection and Model Training

Data sources are grouped into three categories:

Simulation & Synthetic Data : large‑scale 3‑D scenes, simulators, and trajectory synthesis provide supervised navigation samples.

Real‑World & Internet Video Data : capture real robot noise, perception errors, and dynamic environments.

General Multimodal Data : visual‑language corpora supplying semantic priors, reasoning ability, and social norms.

Training strategies are organized into three paths:

Direct Navigation Learning : imitation, sequence prediction, trajectory regression, diffusion generation, or reinforcement learning to learn actions directly.

Auxiliary Task Learning : intermediate goal prediction, sub‑task decomposition, chain‑of‑thought reasoning, future state modeling, map learning, and reward alignment to teach “where to go, why, and what will happen next”.

Vision‑Language Joint Learning : mixing navigation data with general vision‑language data to preserve semantic capabilities and improve instruction understanding, semantic generalization, and cross‑scene transfer.

4. Edge Deployment of Foundation Models

The survey examines deployment on wheeled robots, legged robots, and drones, highlighting platform‑specific constraints. Two acceleration approaches are discussed:

Model‑level structural acceleration : slow‑fast system decomposition, input compression, key‑frame selection, visual token compression, KV‑cache optimization to reduce long‑context inference cost.

Software‑level engineering : cloud‑edge collaboration, asynchronous execution, operator fusion, quantization, and pipeline scheduling to achieve low latency and high energy efficiency on heterogeneous hardware.

Co‑design of robot hardware constraints, model architecture, and inference system is argued as essential for moving from offline evaluation to reliable real‑world operation.

5. Benchmarks and Evaluation Metrics

Five core capabilities are identified for assessing foundation‑model‑driven navigation systems:

Natural‑language instruction conversion to temporally consistent actions.

Goal search and semantic localization under partial observability.

Information acquisition and downstream decision making.

Robustness, generalization, and safety in continuously changing environments.

Retention of abilities across different robot morphologies, sensors, and execution conditions.

Four metric groups are analyzed: task completion, trajectory consistency & semantic alignment, robustness/generalization/safety, and real‑time deployment latency.

6. Conclusions and Outlook

Foundation models shift embodied navigation from task‑specific pipelines to unified multimodal decision frameworks, offering stronger semantic understanding, task generalization, and complex reasoning. The authors highlight three future research directions: (1) overcoming data bottlenecks by establishing scaling laws, (2) integrating vision‑language and world‑model capabilities for simultaneous semantic understanding, instruction following, and future state prediction, and (3) expanding benchmark suites to cover open‑vocabulary goals, dynamic environments, social constraints, real‑time latency, and edge deployment requirements.