“The term Embodied AI is having its moment in the sun right now,” wrote MIT professor emeritus Rodney Brooks in a recent post, “Alan Turing on Embodied Intelligence.” Brooks was a former director of the MIT AI Lab, and the founding director of MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL). In addition, he has founded a number of robotics companies, including iRobot, Rethink Robotics, and Robust.AI.
Brooks has spent much of his career advancing what he calls embodied intelligence — the idea that true intelligence emerges from interaction with the physical world. His post revisited a paper he had written in May 2011, “The Case for Embodied Intelligence,” originally intended for a 2012 conference marking the centenary of Alan Turing’s birth. He included the paper in full, noting that its central ideas — long overshadowed by disembodied, symbolic approaches to AI — are once again attracting serious attention.
Alan Turing—the renowned English mathematician, computer scientist, and cryptologist—published “Computing Machinery and Intelligence” in 1950, one of the most influential papers in the history of AI. It posed the famous question, “Can machines think?” and introduced the Imitation Game, now known as the Turing test.
Less well known is Turing’s 1948 paper “Intelligent Machinery,” in which he distinguished between two kinds of intelligence: embodied and disembodied. Although written earlier, it was not published for a general audience until 1968, fourteen years after Turing’s death.
I had never heard the term Embodied AI (EAI) before—not surprisingly, since it is closely associated with the use of AI in robots and other physical systems, as well as with the impact of the physical world on the development of human intelligence. These are fascinating topics that I know little about, so I was eager to read Brooks’ posts. This is what I learned.
Brooks argues that the earlier paper is the more profound of the two. In his view, “Intelligent Machinery” contained two critical insights. First, Turing suggested that embodied intelligence — that is, intelligence grounded in physical interaction — offered a more reliable path to thinking machines, even though he set it aside as impractical given the technology of his time. Second, Turing introduced the idea of cultural search — the notion that much of human learning comes from immersion in a social and cultural environment.
In Brooks’ opinion, “Turing’s 1948 paper Intelligent Machinery was more important than his 1950 paper “Computing Machinery and Intelligence.”
“For me, the critical, and new, insights in Intelligent Machinery were twofold,” he explained. “First, Turing made the distinction between embodied and disembodied intelligence. While arguing that building an embodied intelligence would be a ‘sure’ route to producing a thinking machine, he rejected it in favor of disembodied intelligence on the grounds of the technical practicalities of the era. Second, he introduced the notion of ‘cultural search’: that people’s learning largely comes from the culture of other people in which they are immersed.”
“Modern researchers are now seriously investigating the embodied approach to intelligence and have rediscovered the importance of interaction with people as the basis for intelligence. My own work for the last twenty-five years has been based on these two ideas. … It is humbling to read Alan Turing’s papers. He thought of it all. First.”
In his recent post, Brooks noted that “Embodied AI is having its moment in the sun now,” as researchers rediscover the concept. “A Brief History of Embodied Artificial Intelligence, and Its Outlook” is one such article. Published in April 2024 by Shaoshan Liu and Shuang Wu, it examines the history, current state, and future of embodied AI.
The authors define Embodied Artificial Intelligence (EAI) as the integration of AI into physical entities —such as robots — giving them the ability to perceive, learn from, and dynamically interact with their environment. Drawing on theories of embodied cognition, they argue that intelligence arises through continuous feedback between an agent’s body and the physical world. The environment, in this view, is not merely a source of sensory input but an active participant in shaping behavior and cognition.
From this perspective, the authors identify three guiding principles for EAI systems:
- they should not rely on predefined, complex logic to manage specific scenarios;
- they must incorporate evolutionary learning mechanisms that adapt continuously to their operational environments; and
- they should treat the environment as a pivotal influence in shaping not just physical behaviors, but also cognitive structures.
While research advances have satisfied these principles individually, the authors note there is no fully functioning commercial system that incorporates all three. When such systems emerge, a major challenge will be enabling them to understand physical laws so they can operate smoothly in the physical world.
They note that many of these ideas trace back to Rodney Brooks’ 1991 paper “Intelligence Without Representation,” where he challenged the dominant AI paradigm of the time by arguing that intelligent behavior could emerge directly from an agent’s physical interaction with its environment without elaborate internal models or symbolic representations.
“Intelligence Without Representation” is a remarkable paper, especially considering that it was written almost four decades ago. Brooks begins by recalling AI’s original ambition: to reproduce human-level intelligence in machines. That ambition was explicit in the 1956 Dartmouth study proposal that coined the term artificial intelligence, which envisioned the development of machines that can in principle simulate every aspect of human intelligence including “language, form abstractions and concepts, solve the kinds of problems now reserved for humans, and improve themselves.”
By 1991, however, Brooks observed that those early hopes had faded. Rather than pursuing general intelligence, the field had retreated into narrower subproblems — knowledge representation, natural language understanding, vision, and other specialized domains.
He reminded readers that we already possess an existence proof of intelligence: humans — and, to varying degrees, many other animals. Drawing inspiration from the 4.6-billion-year evolutionary history of intelligence, he argued for a fundamentally different approach to AI: incrementally build up the capabilities of intelligent systems, and let them loose in the real world with real sensing and real action.
Evolution, Brooks noted, devoted immense time to developing mobility, perception, and survival skills long before language or abstract reasoning emerged. “The first primates appeared 120 million years ago, and the immediate predecessors to the great apes a mere 18 million years ago. Humans arrived in roughly their present form 2.5 million years ago. Agriculture was invented a mere 19,000 years ago, writing less than 5,000 years ago, and ‘expert’ knowledge only over the last few hundred years.”
From this perspective, problem-solving, expert knowledge, and even reasoning may be comparatively easy once the harder problem — effective interaction with a dynamic environment—has been solved.
This is where embodied AI offers a useful counterpoint to today’s AI conversation. The current wave of generative AI has achieved remarkable results by scaling data, computation, and statistical learning — largely in disembodied form. But Brooks’s work, echoing Turing’s earlier insights, suggests that intelligence grounded solely in symbols and text may remain fundamentally incomplete.
If embodied AI is truly having its moment in the sun, it may signal a broader realization: that the next advances in artificial intelligence will not come only from bigger models and more data, but from systems that can move, perceive, act, and learn in the physical world we inhabit.