Introduction to Embodied AI and World Models

Introduction to Embodied AI and World Models

Consider the simple act of picking up a cup of coffee. For a human being, this task requires virtually no conscious effort—we see the cup, reach toward it, adjust our grip based on its weight and the smoothness of its surface, and lift it without spilling. Yet this deceptively simple action involves a remarkable symphony of perception, prediction, motor control, and real-time adaptation. The human brain maintains an implicit understanding of physics—the cup will fall if released, liquids will slosh if moved too quickly, the table will support the cup's weight. It predicts the consequences of actions before executing them, updates its model based on sensory feedback, and seamlessly switches between automatic and conscious control. Now consider asking a robot to perform this same task: the robot must perceive the cup's position and orientation, plan a trajectory to reach it, control its gripper to apply appropriate force, adjust in real-time to variations in cup shape and weight, and execute the entire sequence without human intervention. The gap between human and machine performance on such tasks reveals a fundamental challenge in artificial intelligence—understanding the physical world through bodily interaction.

This chapter addresses the central question driving the field of embodied artificial intelligence: Why do we need AI systems that understand the physical world through bodily interaction, and what role do world models play in enabling such understanding? The answer lies in recognizing that intelligence is not merely the manipulation of symbols or the classification of images in isolation—it emerges from the dynamic interplay between an agent, its body, and a physical environment. By examining the historical evolution of AI approaches, defining the key concepts of embodied AI and world models, exploring the theoretical foundations of the embodiment hypothesis, and surveying practical applications, this chapter establishes the conceptual foundation for everything that follows in this book.

From Symbols to Bodies: The Historical Evolution of AI Approaches

The history of artificial intelligence can be understood as a gradual recognition that intelligence cannot be divorced from physical experience. Understanding this trajectory illuminates why embodied AI represents not merely a technical refinement but a fundamental paradigm shift in how we conceive of intelligent systems.

The Symbolic Era and the Disembodied Mind

The field of artificial intelligence emerged in the mid-twentieth century with an ambitious premise: that intelligence could be captured through the manipulation of symbols according to formal rules. John McCarthy, who coined the term "artificial intelligence" in 1956, envisioned creating machines that could reason logically about the world by representing knowledge as symbolic structures and applying inference rules to derive new knowledge. This approach, known as symbolic AI or Good Old-Fashioned AI (GOFAI), dominated the field for several decades and produced remarkable achievements—theorem provers, expert systems, natural language processors capable of parsing and generating text.

The physical world, in this paradigm, was represented as a set of symbolic facts: `on(cup, table)`, `color(cup, white)`, `material(cup, ceramic)`. Actions were modeled as logical transformations of these facts: `pickup(agent, cup)` would transform the state from `on(cup, table)` to `held(agent, cup)`. The agent itself was conceived as a reasoning engine operating on these symbols, entirely divorced from any physical body or sensory apparatus.

This disembodied approach suffered from a fundamental brittleness. Symbolic representations proved difficult to acquire automatically—knowledge engineers had to manually encode the thousands of facts and rules that humans learn through physical experience. More critically, the symbolic approach struggled with the inherent noisiness and complexity of real-world perception. A cup does not arrive at a vision sensor bearing a neatly formatted `object(cup, position_XYZ, orientation_RPY)` label. Instead, the robot receives a stream of pixel values that must somehow be interpreted as a cup, and this interpretation must remain robust despite variations in lighting, viewing angle, occlusion, and a million other factors that never appear in the neat symbolic world.

By the 1980s, the limitations of purely symbolic approaches had become undeniable. Expert systems, which attempted to capture human expertise in symbolic rules, failed to generalize beyond their narrow domains. Robots programmed with explicit symbolic plans could not cope with the unexpected variations that occur constantly in physical environments. The field began to search for alternatives.

The Rise of Statistical Learning and Neural Networks

The response to symbolic AI's limitations came from an unexpected direction: statistical learning. Rather than hand-crafting symbolic representations, researchers developed algorithms that could learn representations directly from data. The perceptron, invented by Frank Rosenblatt in 1957, had shown that simple networks could learn to classify patterns, but the limitations of single-layer networks seemed to doom this approach. The rediscovery of backpropagation in the 1980s, combined with increased computational power and the availability of larger datasets, enabled the training of multi-layer neural networks that could learn hierarchical representations.

This shift had profound implications for robotics and embodied AI. Rather than requiring engineers to explicitly program every contingency, robots could learn to perceive and act from experience. A robot learning to recognize objects could be trained on thousands of images; a robot learning to navigate could learn from countless traversals. The focus shifted from engineering knowledge representations to engineering learning systems.

However, early neural network approaches in robotics still operated largely in a disembodied fashion. A robot might use a neural network to classify camera images, but the classification was then passed to a separate planning system that operated according to different principles. The neural network was a perception module feeding symbols (or numerical abstractions) to a deliberative system. The body was treated as an output device, not as fundamentally shaping the nature of intelligence.

The Embodied Cognition Revolution

A parallel intellectual movement was occurring in cognitive science and philosophy of mind. Researchers began to argue that human intelligence could not be understood apart from the body that implements it. Workers in embodied cognition, including Rodney Brooks, Andy Clark, and Francisco Varela, proposed that concepts, categories, and even high-level reasoning are fundamentally shaped by the physical properties of the body and its interactions with the environment.

Rodney Brooks, at the MIT Artificial Intelligence Laboratory, became one of the most influential figures in embodied AI. In 1991, he published "Intelligence without Representation," a provocative paper arguing that the project of building intelligent machines should abandon the idea of internal world models and symbolic representations entirely. Brooks's robots used layered finite state machines, each handling a different level of behavioral complexity—from simple obstacle avoidance to more sophisticated navigation behaviors. Each layer processed sensor information and produced motor commands directly, without any central symbolic reasoning. These robots, famously, "had no representation of the world whatsoever." Yet they behaved in surprisingly sophisticated ways, navigating complex environments, avoiding obstacles, and interacting with humans in rudimentary ways.

Brooks's subsumption architecture inspired a generation of researchers and shifted the field's attention toward behavior