Chapter 45: Robots for the Study of Embodied Cognition

The complex study of human intelligence and cognition is moving past the pure computational model, which views cognition as information processing over representations, toward an embodied perspective that mandates incorporating the agent's physical structure and interaction with the real world. Robotics serves as a potent tool for investigating cognition from a "bottom up" approach, starting with basic behaviors. Low-level behaviors, such as walking or obstacle avoidance, often rely primarily on mechanical feedback loops and the interaction between the body and the environment, requiring minimal involvement of cognition. For instance, the passive dynamic walker demonstrates locomotion driven entirely by finely tuned mechanics on a downward slope, without any sensors, motors, or control electronics. The next stage is sensorimotor intelligence, characterized by direct, reflex-like sensorimotor loops, exemplified by simple vehicles developed by Grey Walter and Valentino Braitenberg, which displayed complex patterns through simple, deterministic connections between sensors and motors. Rodney Brooks further developed this concept into behavior-based robotics, proposing a subsumption architecture that explicitly rejected the traditional representational view (GOFAI), arguing that using "the world as its own model" is more effective for simple intelligence. Moving toward minimal embodied cognition involves introducing internal simulation, which extends beyond the "here-and-now" time scale, often taking the form of simple forward models that predict future sensory states based on current state and motor commands. These primitive models allow for rudimentary environmentally decoupled thought and are considered hallmarks of early cognition. The core material for cognition lies in the sensorimotor space, which is critically influenced by the agent's specific embodiment and environmental embedding. Information theory, which quantifies statistical patterns in variables, offers a general method to analyze the structure of information flow, such as how sensory morphology non-linearly transforms environmental input. Crucially, organisms operate in closed sensorimotor loops, necessitating the study of the dynamic relationships between sensors and motors. These relationships are formalized as Sensorimotor Contingencies (SMCs), which are the structured rules governing sensory changes produced by motor actions and are viewed as fundamental building blocks for perception and cognition. SMCs can be classified as modality-related, pertaining to the immediate sensory effects of action based on morphology, or object-related. Using the quadruped robot Puppy and information theory's measure of "transfer entropy" (which quantifies directed information flow between time series), researchers demonstrated how the robot's embodiment creates distinct sensorimotor structures (SM environment). This information flow is then powerfully modulated by specific motor patterns or gaits (SM coordination/strategy), and subtly affected by changes in the environment, such as ground material. Results confirmed the theory that incorporating the action context (gait) significantly enhances the robot's ability to discriminate between different terrains. To study human-like cognition, humanoid robots like the musculoskeletal Roboy and Kenshiro, and developmental robots such as the iCub and fetal simulators, are employed, despite the necessary abstraction from true human physiology. Unlike traditional robotics models, which are explicit, fixed, centralized, and unimodal, biological body representations are plastic, distributed, multimodal, and implicit. Developing brain-like models for robots confers desirable properties like autonomy, robustness, and safety, enabling them to dynamically adapt to changes and construct a "margin of safety" around themselves. Ultimately, robotics supports the embodied and pragmatic turn in cognitive science, providing a synthetic methodology for understanding the brain-body-environment coupling. The highest goal is achieving enactive robotics, which demands that the artificial systems must regulate their sensorimotor interaction in relation to an intrinsic viability criterion—like avoiding battery loss or overheating—in order to generate their own systemic identity and meaning.