Disembodied AI

AI with no body, processing information purely in the digital world — chatbots, image classifiers, and so on.

Disembodied AI is the counterpart to embodied AI: AI with no physical body that only perceives and decides within cyberspace. Its inputs are data such as text, images, and video, and its outputs are text, labels, or images — it never has to act to change the physical environment itself. Large language models like ChatGPT, facial recognition, and recommendation systems all fall into this category. A 2024 embodied-AI survey by Yang Liu and colleagues at Sun Yat-sen University uses a table to contrast the two: in disembodied AI, cognition is separate from any physical entity, while in embodied AI, cognition is fused into an entity such as a robot or a car. The distinction matters because disembodied models learn very strong knowledge from internet data but lack feedback from interacting with the real world; putting one into a robot still leaves problems to solve, such as language grounding, mapping words to real objects and actions, along with producing actual motor output and running in real time.

ExampleChatGPT can write out detailed steps for folding clothes, but it has no hands and cannot actually fold them; getting a robot to carry out those steps is exactly what embodied AI has to solve.

Also called

Non-embodied AI

Related

Embodied AI · Embodied Cognition · Large Language Model · Moravec's Paradox · Language Grounding

Sources

Aligning Cyber Space with Physical World: A Comprehensive Survey on Embodied AI (arXiv 2407.06886)