Most of us are now familiar with AI systems like ChatGPT or DALL-E that learn from vast datasets of text and images. These systems have made remarkable progress in generating content, answering questions, and recognising patterns. However, as the demand for AI systems capable of operating in real-world environments grows, a new frontier in AI emerges: embodied AI—intelligence that learns through physical interaction with the real world.
This article explores how embodied AI transforms robotics by enabling machines to learn and adapt through their bodies and environments. Beyond its immediate applications in robotics, embodied AI also holds potential as a stepping stone toward Artificial General Intelligence (AGI), offering insights into how intelligence grounded in physical experience could bridge the gap between narrow, task-specific AI and more general, adaptable systems.
From Embodied Cognition to Embodied AI
Embodied cognition is the idea that intelligence is not solely a function of the brain but emerges from the interaction between the mind, body, and environment. For instance, when learning to ride a bicycle, we don’t rely on abstract calculations or memorised rules; instead, we develop understanding through physical practice, balance, and sensory feedback. This concept underscores how deeply our thinking and learning are tied to bodily experiences and engagement with the physical world.
Embodied AI applies this principle to artificial intelligence by enabling systems to learn by actively interacting with their environments. Unlike disembodied AI systems, such as Large Language Models (LLMs), which process abstract data like text or images in static, controlled settings, embodied AI agents use sensorimotor coupling and real-world feedback to adapt and improve. These agents act on their surroundings, sense the outcomes of their actions, and adjust their behaviour accordingly. This approach, rooted in embodied cognition, highlights that intelligence emerges from dynamic, physical interactions with the world rather than isolated computation.
The Physical World's Unique Challenges
The real world operates through tangible factors like gravity, friction, balance, and spatial relationships, making it fundamentally different from the abstract nature of digital data. Conventional AI systems, designed to process abstract information, lack the frameworks for agents to interact with their surroundings physically.
Physical learning presents several distinct challenges:
Safety first: Unlike digital mistakes that can be easily erased, physical actions have real consequences. Robots must learn without damaging themselves or their surroundings.
Messy, unpredictable environments: Unlike digital datasets, the real world isn't neatly labelled. Robots must make sense of complex, constantly changing environments with incomplete information.
Learning on the fly: Robots can't pause the world to think. They must process information and adapt in real-time as conditions change around them.
Shifting goals: As tasks and environments evolve, robots must be flexible enough to adjust their strategies and apply previous learning to new situations.
The body matters: A robot's physical design—its sensors, joints, and movement capabilities—directly shapes what and how it can learn about the world. A robot with wheels will learn and interact differently than one with legs or arms.
Current Robots: Specialised but Limited
Today, the most widely adopted robotics solutions rely on precise programming and operate in controlled environments. These robots are commonly found in industries like manufacturing and logistics, where they efficiently perform repetitive tasks such as assembling parts or sorting packages. However, their rigid programming means they lack the flexibility to adapt to changes or unexpected situations. For safety reasons, these robots are typically isolated from humans, as their inability to respond dynamically to their surroundings makes them potentially dangerous in shared spaces. Many systems require human supervision for setup, monitoring, and troubleshooting. While effective for specific, well-defined tasks, this approach limits their ability to function in more dynamic, unpredictable environments.
How Body Information Enhances Robot Learning
Unlike rigid, pre-programmed systems, researchers have been exploring robots that learn from their bodies and environments for decades. Their roots are in cybernetics, which dates back to the 1940s. These robots adapt to the challenges of the physical world by leveraging interaction and feedback, enabling them to function in dynamic and unpredictable settings.
Tactile sensing enables robots to "feel" objects, using touch sensors to gather precise physical feedback. Soft robots, made from flexible materials, adapt their shape and learn through physical deformation. Robots that learn from demonstration, such as those developed at the Toyota Research Institute, acquire a wide range of complex skills by observing human-teleoperated demonstrations and generating actions based on sensor observations. Collaborative robots, like the KUKA LBR iiwa and LBR iisy, operate safely alongside humans in shared workspaces, enabling direct interaction and learning from these experiences to continuously enhance their performance.
Recent advancements in humanoid robotics further demonstrate the potential of embodied learning. These robots are designed to navigate complex environments and perform intricate tasks by learning through physical interaction, making them more adaptable and capable of functioning in real-world scenarios. Together, these developments mark a significant shift toward more flexible, intelligent systems that address the limitations of traditional robotics.
A robot at the Toyota Research Institute learns a new task from a human-teleoperated demonstration. Photograph: Toyota Research Institute https://www.tri.global/news/robots-can-learn-new-actions-faster-thanks-ai-techniques
Embodied AI – a step toward AGI?
Embodied AI is important for robotics and a step toward Artificial General Intelligence (AGI). While large language models (LLMS) are powerful at processing information and simulating reasoning, they lack direct perception and the ability to act in the real world. In contrast, embodied AI grounds intelligence in physical interaction, allowing agents to perceive their environment, take actions, and learn from real-world feedback.
Continuous and dynamic interaction with the real world enables embodied agents to develop a deeper, grounded understanding, grasp causality, and learn more similarly to biological intelligence, as seen in humans and animals. Instead of relying solely on static datasets, embodied AI learns through active engagement, which leads to more flexible, adaptive, and general learning.
Looking Ahead: The Future of Embodied Intelligence
While embodied AI faces significant challenges, such as safety, adaptability, and real-time learning, its potential to transform robotics and move us closer to AGI is undeniable. Despite this promise, embodied AI remains significantly underexplored compared to digital AI systems, which have received enormous attention and investment. However, this gap presents an exciting opportunity to expand the boundaries of AI by focusing on systems that learn through physical interaction. Robots adapting to real-world environments could revolutionise fields such as healthcare, where they could help patients move more easily, or disaster response, by efficiently navigating unpredictable situations.
Embodied AI paves the way for machines to adapt, learn, and evolve in real-world settings. This approach goes beyond advancing robotics—it also moves us closer to the goal of Artificial General Intelligence (AGI). By grounding intelligence in physical experience, embodied AI bridges the gap between narrow, task-specific systems and adaptable, general-purpose intelligence. As we continue to push the boundaries of AI, greater exploration of embodied approaches offers immense potential to create machines that truly understand and navigate the physical world we inhabit.
Opinion article by:
Weronika Wojtak, Researcher in Human-Technology Interaction and Robotics (HTIR)
References
Roy, Nicholas, et al. "From machine learning to robotics: Challenges and opportunities for embodied intelligence." arXiv preprint arXiv:2110.15245 (2021).
Paolo, Giuseppe, Jonas Gonzalez-Billandon, and Balázs Kégl. "A call for embodied AI." arXiv preprint arXiv:2402.03824 (2024).
Foglia, Lucia, and Robert A. Wilson. "Embodied cognition." Wiley Interdisciplinary Reviews: Cognitive Science 4.3 (2013): 319-325.
