cs.AIAug 20, 2026
SaveTowards general embodied intelligence: integrating large language models, knowledge bases, and reasoning capabilities to build the next generation of AI agents
Organizations: College of Mechanical Engineering, Chongqing University of Technology, Chongqing 400054, China · James Watt School of Engineering, University of Glasgow, G12 8QQ, UK · School of Energy and Power, Jiangsu University of Science and Technology, Zhenjiang 212100, China · Magnesium Research Center, Kumamoto University, Kumamoto 860-8555, Japan ·
Abstract
The convergence of large language models (LLMs), structured knowledge bases (KBs), and reasoning ability (RA) presents a promising trajectory toward general embodied intelligence (GEI). This paper reviews the evolution of LLM-centered intelligent systems, emphasising their integration with knowledge representation, logical reasoning, and physical embodiment. We analyse LLM architectures, pre-training methods, and inference mechanisms, along with their interaction with external knowledge sources and structured reasoning frameworks. Furthermore, we examine embodied intelligence (EI) paradigms wherein agents learn and act in physical environments. To synthesise these dimensions, we present a conceptual framework that illustrates the synergy among LLMs, KBs, RA, and embodiment, serving as a guiding model for perception, reasoning, and action rather than an implemented engineering architecture. To advance toward GEI, we identify five key challenges: efficient LLM deployment, closed-loop knowledge integration, hybrid symbolic-neural reasoning, perception-action grounding, and continual learning. This survey provides a comprehensive roadmap for developing adaptive, multimodal agents capable of operating in complex, dynamic settings.
Figures & tables
| Authors | Year | Core Contributions | Limitations |
|---|---|---|---|
| Yao et al. [ 8 ] | 2025 | Reviews multimodal embodied intelligence tasks and models; explores LLM/MLM integration with robots; proposes PPA paradigm and evaluates MLMs. | Classical AI lacks physical interaction; DRL is data-hungry and less robust; VLN training is costly and generalization is poor; LLMs lack causal reasoning; Sim2Real transfer is weak; evaluation standards are lacking. |
| Dong et al. [ 9 ] | 2025 | Analyzes LLM-enabled Human-Robot Symbiotic Manufacturing (HRSM), including interaction, collaboration, and execution. | Privacy and data sensitivity risks; hallucination and bias in LLMs; absence of unified evaluation standards; integration difficulties in diverse environments; token length limits document processing. |
| Tan et al. [ 10 ] | 2025 | Proposes CEIM to fuse multi-source data and LLM reasoning for semantic integration and decision-making in customized manufacturing. | Data fusion issues; low semantic alignment; real-time control accuracy is lacking; high computation cost; insufficient standardization of AI-device interfaces. |
| Jeong et al. [ 11 ] | 2024 | Surveys five directions of LLM/VLM in robotics: reward design, control, planning, manipulation, and scene understanding. | Limited by embedded compute; weak sensor fusion; prompt bias and poor interpretability; lacks domain-specific knowledge; Sim2Real generalization is weak. |
| Wang et al. [ 12 ] | 2025 | Discusses LLMs in robotics tasks and evaluates GPT-4V; explores future applications in agriculture and medicine. | Plan robustness is lacking; long prompts require expert input; fixed action sets reduce flexibility; API is closed-source; RGBD-based 3D reasoning is limited. |
| Our Survey | 2025 | Systematically analyzed the synergy among LLMs, knowledge bases, reasoning capabilities, and embodied intelligence; proposed a unified framework and five future directions for GEI. | Limited empirical examples; insufficient discussion on multilingual settings and real-world deployment. |
TABLE I: Comparative Summary of Related Embodied Intelligence Review Articles
Fig. 1: Schematic diagram of a unified multimodal large language model architecture.
Fig. 2: Evolution and Taxonomy of Large Language Models.
Fig. 3: Schematic architecture of the Transformer model.
Fig. 4: Construction Process of Large Language Models.
Fig. 5: Description of the execution process of the intelligent document processing and knowledge question answering system.
| Author(s) | Core Improvement | Advantages | Application Scenarios |
|---|---|---|---|
| Cummins et al. [ 40 ] | Compiler optimization framework | Enhances disassembly processes | Compiler optimization |
| Hu et al. [ 41 ] | Prompting strategies incorporating context to enhance VQA performance | Strengthens contextual understanding | Knowledge-based applications |
| Hu et al. [ 42 ] | LLM-TIKG: Threat Intelligence Knowledge Graph framework | Enhances cybersecurity analysis | Threat Intelligence Analysis |
| Qian et al. [ 43 ] | Infrastructure for large-scale LLM training | Efficiently handles large-scale training tasks | LLM training infrastructure |
| Sun et al. [ 44 ] | UniCoder: Intermediate representation to improve downstream NLP task processing | Enhances the model’s ability to handle complex tasks | Model scaling |
| Zhang et al. [ 45 ] | BB-GeoGPT: Framework tailored for Geographic Information Science | Highly adaptable to domain-specific knowledge | Geographic Information Science (GIS) |
TABLE II: Comparison of Research on Large Language Models (LLMs)
Fig. 6: Flowchart of knowledge base construction and use in large language model.
Fig. 7: Summary of inference optimization techniques in large language models.
Fig. 8: The core components and development path of embodied intelligence.
Fig. 9: The development of embodied intelligence.
Fig. 10: Different categories of Embodied Robots: (a) Quadruped Robots, (b) Fixed-base Robots, (c) Tracked Robots, (d) Wheeled Robots, and (e) Humanoid Robots.
Fig. 11: Example diagram of the integration of embodied intelligence and large language model.
| Author | Year | Research Topic | Key Contribution | Application Domain |
|---|---|---|---|---|
| Zardini et al. [ 101 ] | 2020 | Co-Design Framework | Structured co-design for integration | General systems |
| Ma et al. [ 102 ] | 2021 | Soft Robotics Actuators | Integrated sensing & actuation | Soft robotics |
| Gupta et al. [ 104 ] | 2021 | DERL for Adaptation | Co-evolution of body and control | Locomotion & control |
| Roy et al. [ 105 ] | 2021 | ML in Embodied Systems | Energy-aware physical learning | Robotics & ML |
| Mengaldo et al. [ 103 ] | 2022 | Physics Modeling | Physics-based modeling | Soft robotics |
| Iida et al. [ 106 ] | 2022 | Temporal Dynamics | Timescale-based self-organization | Adaptive systems |
TABLE III: Summary of Key Contributions in Embodied Intelligence Research
Fig. 12: Embodied Intelligence, Large Language Model, Knowledge Base and Reasoning Ability together constitute the core architecture of the intelligent system.
| Author(s) | Innovation | Advantages | Drawbacks |
|---|---|---|---|
| Daniel et al. [ 111 ] | Verifying RDF triples using LLMs by comparing RDF statements to external documents | High precision (88%) in RDF verification | Requires human oversight, method may be slower |
| Feng et al. [ 112 ] | Knowledge Graph–based thought framework to improve LLM responses with verified KG information | Reduces factual errors, enhances drug–cancer associations | Limited to biomedical, requires cross-domain testing |
| Dai et al. [ 113 ] | Empirical study on LLMs’ comprehension of KG input formats and prompt organization | Improves LLM performance in KG-based tasks | Varying effectiveness across different models, noisy subgraphs |
| Erickson et al. [ 114 ] | Enhancing LLM workflows with RAG and KGs to ensure traceability and flexibility | More accurate and explainable LLM outputs | Integration with external tools can be complex |
| Xiang et al. [ 115 ] | LLM-based generative retrieval framework focusing on trustworthy search results | Improves relevance, responsibility, and trustworthiness | Works better with certain query types, not universally applicable |
| Sequeda et al. [ 116 ] | LLM-powered enterprise question answering, supported by KGs | Provides formal validation and trusted data for queries | May not scale well for large or complex queries in diverse enterprises |
TABLE IV: LLM and Knowledge Graph Integration
| Authors | Innovation | Advantages | Drawbacks |
|---|---|---|---|
| Wang et al. [ 117 ] | LLM-KGMQA: Multi-hop question-answering system for the medical field, addressing entity linking and reasoning | Improved accuracy (99.80%) and robust performance | Requires significant computational resources for entity linking and multi-hop reasoning |
| Jiang et al. [ 118 ] | Knowledge-guided agricultural LLM for domain-specific decision-making | State-of-the-art performance in agriculture, improves accuracy in agricultural queries | Data scarcity in the domain, challenges in precision retrieval |
| Benjira et al. [ 119 ] | LLM-augmented KG for SDG indicator mapping, integrating open data | High precision in mapping open data to SDGs | Requires access to high-quality and diverse metadata |
| Freire et al. [ 120 ] | LLM-powered knowledge-sharing system for manufacturing, retrieving information from factory documentation | Enhances quick information retrieval and resolution of issues | Preference for human experts still prevails, open-source models may have privacy concerns |
| Xie et al. [ 121 ] | Theme-based lecture summary system using LLMs and graph-based segmentation | Enhances knowledge recall and student satisfaction | Effectiveness varies with lecture structure |
| Sun et al. [ 122 ] | SF-GPT: A training-free method for enhancing LLM-based KG construction | Improved recall and F1 score, no training required | May not work well with all types of datasets or knowledge fusion |
TABLE V: LLM for Domain-Specific Applications
| Authors | Year | Innovation | Advantages | Limitations |
|---|---|---|---|---|
| Wu et al. [ 123 ] | 2025 | Proposed Long-to-Short (L2S) reasoning and model merging. | Increases reasoning efficiency while maintaining performance. | Merging efficiency depends on model size, adds computational complexity. |
| Tung et al. [ 124 ] | 2025 | Introduced GreenMind, using strategy optimization for reasoning. | Outperforms previous works on Vietnamese reasoning tasks. | Limited to specific languages, struggles with low-resource languages. |
| Ma et al. [ 125 ] | 2025 | Introduced SQL-R1, a reinforcement learning-based model for NL2SQL. | Improves accuracy in complex database tasks. | Reinforcement learning training may suffer from cold start issues. |
| Costas [ 126 ] | 2024 | GNN-RAG combines GNN reasoning with LLMs in a retrieval-augmented generation style. | Optimizes graph data processing, improves knowledge graph QA. | Performance depends on graph data quality, limited in complex reasoning. |
| Sui et al. [ 127 ] | 2024 | Proposed FiDeLiS for factuality-enhanced LLM reasoning. | Improves factuality and interpretability. | Dependency on reasoning steps increases computational cost. |
| Luo et al. [ 128 ] | 2025 | Introduced RALLRec+ for enhancing recommendation system reasoning. | Enhances recommendation reasoning, handles dynamic user preferences. | Requires continuous data updates, may still be less effective in real-time applications. |
TABLE VI: Category 1: Research Based on Reasoning Frameworks and Methods
| Authors | Year | Innovation | Advantages | Limitations |
|---|---|---|---|---|
| Yang et al. [ 129 ] | 2025 | Introduced Reasoning-based Bias Detector (RBD). | Reduces bias in evaluations, improves consistency. | Needs significant supervision for bias detection. |
| Zhang et al. [ 130 ] | 2024 | Designed step-wise multimodal models for document reasoning. | Enhances multimodal understanding, improves complex question answering. | Relies heavily on high-quality labeled data. |
| Beary et al. [ 131 ] | 2025 | Introduced hybrid fine-tuning to enhance medical domain reasoning. | Improves reasoning accuracy in radiation oncology. | Domain-specific, limited transferability to other medical fields. |
| Tan et al. [ 132 ] | 2025 | Proposed Hydra framework for cross-source enhanced reasoning. | Excels in multi-hop and multi-entity reasoning tasks. | Requires diverse data sources, computationally intensive. |
| Kunat et al. [ 133 ] | 2025 | Merged language-specific LLMs with advanced reasoning models. | Enhances reasoning in low-resource languages. | Can be computationally expensive for merging models. |
| Lan et al. [ 134 ] | 2024 | Proposed LLM4QA for graph-based reasoning with SPARQL queries. | Improves knowledge graph reasoning efficiency. | Relies on structured knowledge graphs, limited in unstructured data. |
TABLE VII: Category 2: Application-Specific Reasoning Research
| Authors | Year | Innovation | Advantages | Limitations |
|---|---|---|---|---|
| Fan et al. [ 138 ] | 2025 | Proposed a three-stage framework for industrial robot task execution using LLMs | Outstanding GPT-4 performance in industrial tasks | Limited ability in 3D spatial planning |
| Chen et al. [ 139 ] | 2025 | Designed RoboGPT with long-term decomposition and feedback adjustment | Adaptive subgoal updating | Requires large datasets and complex modules |
| Ruaridh et al. [ 141 ] | 2025 | Introduced ELLMER with RAG for multi-step reasoning | Strong sensorimotor adaptation | Tested mainly in home-like tasks |
| Li et al. [ 143 ] | 2024 | Proposed multi-modal LLMs for autonomous mining driving | Applicable to realistic mining scenarios | Conceptual work, lacks implementation |
| Song et al. [ 146 ] | 2023 | Introduced LLM-Planner for few-shot planning | Strong performance with limited data | Weak environmental state modeling |
| Andrew et al. [ 147 ] | 2025 | Unified GEA model across embodied tasks via MLLM | High cross-domain generalization | Data-hungry and complex training |
TABLE VIII: LLM-driven Embodied Task Execution and Planning
| Authors | Year | Innovation | Advantages | Limitations |
|---|---|---|---|---|
| Sun et al. [ 140 ] | 2025 | Multi-level LLM framework for coal mine sensor analysis | Efficient risk prediction and learning | Requires simulated and expert knowledge |
| Abhay et al. [ 154 ] | 2024 | EnvGen: LLM-generated adaptive RL training environments | Boosts training efficiency | Depends on feedback quality |
| Zheng et al. [ 155 ] | 2023 | Steve-Eye combines LLM and visual encoders | Enables end-to-end multimodal interaction | Requires large datasets |
| Ma et al. [ 156 ] | 2025 | Perspective on LLMs in autonomous driving | Offers future research outlook | Lacks technical solutions |
| Liu et al. [ 157 ] | 2024 | EAI-SIM simulation platform with ROS | Controls UAVs and arms in photo-realistic sim | Needs powerful simulation hardware |
| Song et al. [ 158 ] | 2024 | Guide-LLM aids visually impaired via text maps | Combines commonsense and path planning | Small sample evaluation |
TABLE IX: Multi-Modal Fusion and Perception Enhancement in Embodied Systems
| Authors | Year | Innovation | Advantages | Limitations |
|---|---|---|---|---|
| Zhang et al. [ 162 ] | 2024 | ReAd mechanism improves multi-agent LLM efficiency | Fewer LLM queries and agent steps | Evaluation mainly in simulation |
| Liu et al. [ 163 ] | 2024 | FaGeL uses smart fabric for implicit feedback | Learns user preference non-intrusively | High system complexity |
| Guo et al. [ 164 ] | 2024 | Organizational prompting reduces LLM agent conflict | Boosts teamwork and efficiency | Validation in real-world pending |
TABLE X: Multi-Agent Collaboration and Organizational Strategies
| Authors | Year | Innovation | Advantages | Limitations |
|---|---|---|---|---|
| Liu et al. [ 142 ] | 2025 | Contextual backdoor attacks in LLM-based agents | Dual-modal triggers for covert actions | No defense strategy yet in real use |
| Zhang et al. [ 144 ] | 2024 | Investigated LLM-based robot jailbreak threats | Systematic exposure of jailbreak scenarios | No concrete safety mechanisms |
| Li et al. [ 145 ] | 2024 | Universal benchmarking interface for LLM agents | Fine-grained error taxonomy | No coverage of real multimodal scenes |
| Kovalev et al. [ 153 ] | 2022 | Reviewed LLM-based instruction planners without training | Works in zero-shot scenarios | Lack of real task validations |
| Szot et al. [ 165 ] | 2023 | LLaRP: RL-trained LLM policy with vision | Superior multi-task performance | Needs intensive training and environment |
| Leon et al. [ 166 ] | 2024 | Studied virtual agent personality effects in VR | Extrovert agents yield better experience | No embodiment in physical robots |
TABLE XI: Security and Robustness Challenges in Embodied LLMs
| Authors | Year | Innovation | Advantages | Limitations |
|---|---|---|---|---|
| Ruaridh et al. [ 168 ] | 2025 | Proposed ELLMER framework combining GPT-4 with RAG for task planning in unpredictable environments. | Enables long-horizon task execution via visual/force feedback; adaptable planning. | Generalization scope not fully validated. |
| Fan et al. [ 169 ] | 2024 | Introduced LLM agents in manufacturing for autonomous design, planning, and control. | Achieved 81.88% task completion with GPT-4 in complex scenarios. | Challenges remain in 3D spatial task handling and real-time integration. |
| Liang et al. [ 170 ] | 2023 | Introduced “Code as Policies” using LLMs to write robot policy code from commands. | Enables reactive and trajectory-based control via few-shot prompting. | Limited physical-world validation. |
| Xu et al. [ 171 ] | 2025 | Reviewed LLM-integrated CPSS frameworks for human-centered Industry 5.0. | Connects LLMs, perception, scheduling, and swarm intelligence into unified framework. | Lacks empirical task-specific evaluations. |
| Zhao et al. [ 172 ] | 2024 | Developed MultiBotGPT system using GPT-3.5 for UAV/UGV task execution. | Outperforms BERT in assignment success rate; improves operator experience. | Focused on limited command types and single-modal input. |
| Shen et al. [ 173 ] | 2024 | Discussed LLMs as foundation for embodied agents in physical environments. | Conceptual bridge from virtual to real-world through embodiment. | Lacks concrete framework or task-based implementation details. |
TABLE XII: Summary of Studies on LLM-driven Embodied Control
More in this topic
Embodied AI
Embodied AI studies agents that perceive, reason, and act through bodies in physical or simulated environments.
See all 255 papers