Physical Intelligence Research Articles

Physical intelligence as a new paradigm

Intelligence of physical agents, such as human-made (e.g., robots, autonomous cars) and biological (e.g., animals, plants) ones, is not only enabled by their computational intelligence (CI) in their brain, but also by their physical intelligence (PI) encoded in their body. Therefore, it is essential to advance the PI of human-made agents as much as possible, in addition to their CI, to operate them in unstructured and complex real-world environments like the biological agents. This article gives a perspective on what PI paradigm is, when PI can be more significant and dominant in physical and biological agents at different length scales and how bioinspired and abstract PI methods can be created in agent bodies. PI paradigm aims to synergize and merge many research fields, such as mechanics, materials science, robotics, mechanical design, fluidics, active matter, biology, self-assembly and collective systems, to enable advanced PI capabilities in human-made agent bodies, comparable to the ones observed in biological organisms. Such capabilities would progress the future robots and other machines beyond what can be realized using the current frameworks.

Self-Organization, Entropy Production, and Physical Intelligence

Presenting thermodynamic systems from the modern thermodynamic point of view, an effort is made to discuss physical intelligence (PI) entirely in terms of thermodynamic forces, flows, stability. The following are noted: (a) A reasonable starting point for inquiry into PI is dissipative structures: ordered processes that require continuous production of entropy for their maintenance. The entropy production rate increases when irreversible processes form intricate networks capable of interacting with the environment in complex ways. High entropy production rates may be a necessary condition for PI. (b) Availability of a large number of dissipative structures, among which a system makes transitions when it interacts with its environment, may constitute an additional necessary condition for PI. Inter-state transitions enhance the rate of entropy production and they could ground a physical perspective on perception-action. (c) PI tied to dissipative structures would imply that it is non-algorithmic. (d) PI variants of “intentionality” and “survival” may be expressible, at least in some cases, in terms of thermodynamic stability, either of a process or a network of processes.

On Intelligence From First Principles: Guidelines for Inquiry Into the Hypothesis of Physical Intelligence (PI)

Gibson's (1966, 1979/1986) ecological approach to perception, action, and cognition is patently non-representational and non-computational. In the place of these commonly assumed neural endowments, ecological psychology seeks to expose the laws that underlie intelligent capabilities. It is argued that this is the goal for understanding the directed behavior of not just humans, not just animals, and not just the living. We argue that an approach to intelligence that is physically grounded is completely consistent with—and is even a natural consequence of—the central tenets of ecological psychology. We identify two dozen guidelines for seeking intelligence in first principles.