Embodiment
Murray Shanahan & Ben Henke
An agent is embodied if it interacts with the world through a body that perceives, acts, and occupies space.
Does an AI system need a body to truly understand and intelligently engage with the world, or can disembodied systems achieve the same level of cognition?
Key Points:
Embodiment refers to an agent’s physical or virtual presence in the world, allowing it to perceive and act through a body.
According to the embodied cognition paradigm, an agent’s physical interactions fundamentally shape its intelligence and cognitive processes.
Embodied AI systems have unique benefits and risks
Embodiment is a critical concept for both biological and artificial agents, as it shapes how an agent interacts with its environment. In AI, understanding the conditions and implications of embodiment can influence how we design, engage with, and interpret these systems. From robots in healthcare to autonomous vehicles, embodied AI systems are becoming more prevalent, raising important questions about intelligence, interaction, and the human tendency to anthropomorphize machines.
What is Embodiment?
An embodied agent is one that interacts with its environment through a body that serves as the center for its sensory perceptions and actions. This body provides a physical presence, occupies space, and allows the agent to interact directly with objects and other agents. Humans and animals are naturally embodied, using their bodies to perceive the world through senses like sight, touch, and hearing, and to act upon it. In AI, embodied agents include robots like household assistants that can navigate rooms, pick up objects, and respond to human gestures. Autonomous drones that fly and collect data, or self-driving cars that transport passengers, are also examples of embodied AI systems. These agents’ bodies have sensors and effectors that enable them to perceive their surroundings and act accordingly. However, not all systems that interact with the world count as embodied. For instance, a robotic arm controlled by a human operator does not qualify as an embodied agent, as it lacks autonomous agency. Similarly, chatbots are not embodied because they neither occupy space nor have a physical location.
Vague Cases
While there are clear examples of embodied artificial agents, some cases are more ambiguous. For instance, in video games, non-playable characters can be virtually embodied—occupying a virtual space—without being physically embodied. Virtual embodiment may carry some of the implications of physical embodiment, such as influencing cognitive processes, but lack others, such as raising ethical concerns tied to physically embodied agents.
Another key question is the extent to which a system must engage with its environment to count as embodied. While humanoid robots are clearly embodied, other systems, such as robotic cameras tracking moving objects, autonomous vehicles, or drones are more ambiguous. These systems may lack full agency, a fully developed body, or both. More abstract cases include AI systems that control multiple independent robots or systems (like chatbots) that interact through digital mediums (e.g., text outputs or pixels on a screen). These cases may exhibit some features of embodiment but not all.
Embodied Cognition
Embodied cognition is the theory that an agent’s cognitive processes are deeply rooted in its physical interactions with the environment. According to this view, intelligence is not just a product of abstract computations in the brain but emerges from the dynamic relationship between the body, the mind, and the world. For example, a robot learning to walk must coordinate its motors, balance, and sensory feedback to move effectively. This physical experience shapes its “understanding” of movement in a way that a simulated system would not replicate. Embodied cognition challenges traditional models that separate mind and body, suggesting that to achieve human-like intelligence, AI systems may need to be embodied. For some thinkers — who draw a tight connection between embodiment and the maintenance of a metabolism, a stronger notion than that discussed above — genuine embodiment may be beyond the reach of current or near-future AI systems.
Benefits and Risks of Embodiment
Embodied AI systems present unique benefits and risks. Embodiment can obviously enable an AI system to perform tasks not possible for an non-embodied system. However, physical agents can also cause physical harm, such as when a driverless car crashes. Moreover, embodied systems, especially those designed to resemble humans or animals, are more likely to be anthropomorphised. This tendency can lead to risks, such as users mistakenly attributing intentions or consciousness to the system. On the other hand, anthropomorphization could also be beneficial, especially, but controversially, in therapeutic contexts, where robots designed to appear more human-like may improve social interactions and care delivery.