Embodied AI

What it means

Embodied AI refers to AI systems that act through a physical body and learn from the results of their actions. A 2025 arXiv paper on multi-agent embodied AI describes these systems as combining AI with physical bodies so that they can perceive, reason and interact with their environments, using sensors for input and actuators for action.

The idea has roots in cognitive science. Embodied cognition theories investigate how cognition is shaped by the bodily state and capacities of an organism, and embodied AI applies a similar view to machines.

How it applies to humanoid robots

Wikipedia defines a humanoid robot as one resembling the human body in shape, which makes it one platform on which embodied AI can be tested. Embodied AI is not limited to humanoids: wheeled robots and arm-only systems are embodied too.

Figure's Helix announcement is an example of an embodied model applied to a humanoid. It describes a vision-language-action model that controls the whole upper body at high rate. The post is a company announcement, so its performance statements are the company's own.

Physical AI as an alias

IBM describes physical AI as AI systems that operate in and interact with the physical world, typically combining AI models with sensors, actuators and other control systems.

IBM contrasts this with robotics, which it describes as focused on the mechanics and control of physical machines. Usage varies between writers, so this site uses embodied AI as the umbrella term and treats physical AI as an alias rather than a separate category.

Counting and terminology caveats

Embodied does not mean humanoid. A wheeled robot or an arm with a learned model is also embodied, which is why the site records the form factor separately from the AI label.

A named model is not the category. Names such as Helix describe specific models; embodied AI is the broader category for systems in which a model's decisions act through a body.

Demonstrations are not deployments. A company demo can show a task done once, so the site records such claims with their source and confidence rather than as a general capability.

How to read embodied AI on humanoidversus.com

This entry has no live table. Use the ai_stack and actuation fields on each robot page, and check the confidence level and note before comparing claims.

For further reading, see the 2025 arXiv paper by Feng et al., 'Multi-agent Embodied AI: Advances and Future Directions', IBM's physical AI explainer, Wikipedia's article on embodied cognition, and Figure's Helix announcement.