Physical intelligence as a new paradigm

Abstract

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.

Keywords: Physical Intelligence; mechanical computation; mechanical memory; mechanics; meta materials; multistability.

Fig. 1

Physical intelligence (PI) components and enablers on a physical agent body interacting with its environment. Such body PI is also coupled to the computational intelligence (CI) in the brain, where the embodied intelligence (EI) field investigates the tight coupling between the body and brain.

Fig. 2. Autonomous human-made machine examples using mainly physical intelligence.

(a) Theo Jansen’s human-size kinetic art machines called strandbeest [12] (Animaris Currens Ventosa 1993, photo: Adriaan Kok) using self-powering/actuation using wind energy harvesting-based motion generation, mechanical sensing and self-regulation for avoiding the sea water, and walking multi-legged mechanisms and feet optimized for a robust and efficient walking on sandy beaches. (b) Self-regulating mechanical control system example from the history: centrifugal governors were used as self-regulating mechanical control systems in James Watt’s steam engines from 1788 (Adapted from [13]. Copyright, George Routledge and Sons). Similar self-regulating mechanical control systems have been implemented these days in some vehicle transmissions and record players. (c) A soft pneumatic logic circuit and memory, outputting a binary value of pressure depending on the most recent non-zero input signal of pressure (Aapted from [14]. Copyright 2019, National Academy of Sciences).

Fig. 3

Inspired by biological microhair structures, diverse multifunctionalities can be encoded to the agent body surfaces using synthetic microfiber arrays: surface adhesion and friction of vertical or angled microfiber structures can be controlled using active or passive, vertical or lateral mechanical load control; microfiber arrays can be designed to super-repel or super-attract specific or all liquids in the operation environment; fibers can reduce the fluid or air flow-induced drag forces on the body; they can control the heat and electrical conductance of the agent surface with or without contacting to another surface; they can be used as a flow or contact sensor if a transduction material is integrated to the fiber base or structure; and they can induce ciliary motion-based fluid or solid object transport if they are made of an active material.

Fig. 4. Physical intelligence-based learning and decision-making examples in biological agents.

(a) Stentor roeseli single-cell organism has a form of sequential decision-making to avoid irritating repeated stimuli (Adapted with permission from [107]. Image credit: Joseph Dexter and Sudhakaran Prabakaran. Copyright 2019, Current Biology). (b) A slime mold (yellow Physarum polycephalum) can learn and solve mazes and the traveling salesperson problem in an energy efficient way if a food attractant is placed at target locations (Adapted with permission from [108]. Copyright 2010, Science, AAAS). (c) The hunting cycle of the carnivorous Venus flytraps (Dionaea muscipula) (photo credit: Noah Elhart [109]) involves multiple decision-making steps to trap, fully close/seal and digest a landed prey insect.. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)