Overview
We are building a theoretical framework that will guide the creation of artificial agents that adjust their neural networks (brains) to feedback from their bodies and surroundings -- in essence to learn how to navigate their surroundings.
Humans are not born pre-programmed with a set of intelligent behaviors such as carrying a conversation or moving through a crowded room. Rather, we build this kind of intelligence by adjusting to constant feedback from our own bodies and surroundings.
Artificial intelligence (AI)—the kind that would allow future robots to behave like natural organisms—should arise the same way. Although the field of AI is currently dominated by the pre-programming paradigm, where the robot is instructed to perform specific behaviors, we believe the notion of embodiment can help us evolve truly intelligent artificial systems. In the field of embodied intelligence, this notion places emphasis on the role of an agent’s body in generating behavior. Isolated examples of artificial embodied agents exist, but there is no quantitative theory for their design.
Our project aims to build a theoretical framework for embodied artificial intelligence that treats the brain, body, and environment as inseparable cognitive elements. This theory will guide the creation of artificial agents that adjust their neural networks (brains) to feedback from their bodies and surroundings. In our experiments, we do not program robots with specific behaviors such as taking steps. Rather, we tell them to seek out a rich variety of experiences by instructing them to maximize their “predictive information,” a notion developed within complexity science. From this principle, the agents’ behaviors emerge in a self-organized way. Like human children, our embodied agents discover their own behavioral possibilities through an interactive and exploratory process.
As a consequence of this learning process, an agent’s body and environment are heavily involved in controlling its behavioral modes, thereby reducing the amount of control required from the brain. This is referred to as “cheap design,” and is considered to be one of the principles of embodied intelligence.
As we devise a quantitative and systematic theory for designing these agents, we hope to build our own understanding of the complex interplay between body, environment, and brain.
The Santa Fe Institute, known for its pioneering work in information theory and self-organization, is an ideal setting to conduct our research. As we put concepts from information geometry toward novel applications in the fields of robotics and AI, our aim is to create intelligent embodied agents that can adapt to their own bodies and environments.