Written and maintained by CASRAI Editorial Board
Last updated
Embodied AI is artificial intelligence built around an agent that has a body, physical or simulated, and learns by perceiving and acting in an environment rather than only processing static data. This guide is for researchers and research administrators who need a working definition, the components usually named, the main research areas, and a way to navigate the survey literature. The field has no single fixed formal definition, so the sections below attribute each framing to the paper that states it (arXiv pages read 2026-11-07).
What embodied AI means
In the ICML 2024 position paper A call for embodied AI (Paolo, Gonzalez-Billandon and Kégl), the authors propose embodied AI as “the next fundamental step in the pursuit of Artificial General Intelligence” and contrast it with current advances in large language models (LLMs). They trace the idea of embodiment through philosophy, psychology, neuroscience and robotics. In short: an embodied AI system is judged by what it does in an environment, not only by the text or images it produces.
The survey Aligning Cyber Space with Physical World: A Comprehensive Survey on Embodied AI (Liu et al., arXiv:2407.06886) describes the field as covering both physical robots and simulators, and likewise treats embodied AI as important for progress toward artificial general intelligence. For the wider discipline, see What Is Artificial Intelligence?
Core components of an embodied agent
The position paper names four essential components of an embodied agent: perception, action, memory and learning. Its proposed framework is built from cognitive architectures and aligned with Karl Friston’s active inference principle. Read as a loop, perception takes in sensor information, action changes the environment, memory retains what happened, and learning updates behaviour from the outcome.
Research areas
The comprehensive survey (arXiv:2407.06886, first submitted 9 July 2024, revised 25 August 2025, subject class cs.CV) organises the field around four research targets:
| Research target | What it is about |
|---|---|
| Embodied perception | Understanding the environment from sensor input |
| Embodied interaction | Acting on and responding to the environment |
| Embodied agents | Systems that combine perception and action to pursue goals |
| Sim-to-real adaptation | Transferring what is learned in simulation to the physical world |
The same survey also discusses the role of multi-modal large models and world models, plus datasets and open challenges. The four target names come from the survey’s abstract; the one-line descriptions are editorial plain-language glosses.
Note how the two sources divide the topic differently: the position paper lists the components inside an agent (perception, action, memory, learning), while the survey lists research targets across the field. They are two views of the same subject, not competing definitions.
Embodied AI compared with LLM-only and static-learning systems
| Dimension | Embodied agent | LLM-only or static-learning system |
|---|---|---|
| Environment | Perceives and acts in a physical or simulated environment | Learns from a fixed collection of data |
| Components emphasised by Paolo et al. | Perception, action, memory and learning together | Position paper contrasts embodied AI with these approaches as the next step beyond them |
This comparison follows the framing of the position paper’s abstract. It does not say that LLMs are unimportant to embodied AI: the comprehensive survey explicitly covers multi-modal large models as part of the field. For how language models are trained to follow human preferences, see RLHF.
Reading the literature
- Start with the position paper (arXiv:2402.03824, submitted 6 February 2024, version 4 dated 13 September 2024, ICML 2024 Position Paper track) for the conceptual argument and the four components.
- Use the comprehensive survey (arXiv:2407.06886) for the map of perception, interaction, agents and sim-to-real work, datasets and challenges.
- Check the version. Both papers have several arXiv versions; cite the version you read. Many arXiv papers are preprints that may change after peer review.
FAQ
What is embodied AI?
AI in which an agent with a physical or simulated body learns and acts through interaction with its environment. Paolo et al. (2024) name perception, action, memory and learning as its essential components.
Is embodied AI the same as robotics?
Not exactly. The comprehensive survey covers both robots and simulators, so embodied AI is a research field about agents that perceive and act, which includes but is not limited to physical robots.
How does embodied AI relate to LLMs?
The position paper contrasts embodied AI with advances in LLMs, while the comprehensive survey treats multi-modal large models as one of the technologies used in embodied systems.
Is there one agreed definition?
The two sources read here do not give one formal definition in their abstracts; each frames the field through its own components or research targets.
Sources
- Paolo, G., Gonzalez-Billandon, J., Kégl, B. A call for embodied AI. ICML 2024, PMLR 235:39493-39508. arXiv:2402.03824, https://arxiv.org/abs/2402.03824 (read 2026-11-07).
- Liu, Chen, Bai, Liang, Li, Gao, Lin. Aligning Cyber Space with Physical World: A Comprehensive Survey on Embodied AI. arXiv:2407.06886, https://arxiv.org/abs/2407.06886 (read 2026-11-07).
