Embodied endosymbiosis for Protean machines
Abstract
Living systems are unparalleled in their ability to adapt to new circumstances. One of the ways they accomplish this is through aggregation: if A and B cannot solve a new challenge on their own, they may by joining into AB. This aggregation has occurred at different levels of organization at different times during evolution. Early on, one cell was engulfed by another, and the conjoined pair eventually evolved into eukaryotic cells. The engulfed cells evolved into mitochondria, which provide a service to their hosts: energy production. This process of endosymbiosis occurs at the organismal level as well: bacteria in the mammalian gut provide important digestive functions while being rewarded with a safe and nutrition-rich environment. Finally, humans aggregate into societies, and human societies are in the process of fusing with their increasingly capable technological artefacts. Biological evolution thus demonstrates that the path to complex, general systems may be more readily attainable by training and then fusing simple and specialized systems into aggregates, and then aggregates of aggregates, rather than continuously reprogramming a monolithic, unchanging physical system. In the technical disciplines, investigators often follow this latter path. This causes most machines to forget old tasks as they learn new tasks, an unsolved problem known as catastrophic forgetting [22]. Similarly, state-of-the-art robots are still only able to exhibit very specialized behaviors in controlled environments. Even among investigators who attempt to imbue machines with morphological computation, they often design them with fixed body plans: The bodies may be capable of complex movements with simple controllers, but they cannot absorb other machines and inherit their capabilities [56, 15, 19]. Similarly, soft machines can change shape [71, 9, 66, 38], but they cannot absorb other robots into themselves and retain the behaviors both were capable of performing alone. Finally, modular robots can combine [83, 62, 41], but it has never been demonstrated that they can do so such that they accrue new behaviors without sacrificing their specialized behaviors. Thus, we here propose an evolutionary algorithm that relaxes a central assumption in evolutionary design applied to embodied machines: machines cannot absorb other machines into themselves. More specifically, we propose to evolve not just the body plans and control policies of simple specialists, but also the way in which they aggregate. The fitness of any one aggregate will be its ability to retain the specialists behaviors while also achieving new behaviors. We refer to this as embodied endosymbiosis: machines physically combine into new shapes and topologies, with some of the original machines being partly or completely absorbed into other machines. To achieve this we will first develop a co-evolutionary algorithm composed of two populations. Control policies in the first population attempt to generate desired behavior in all aggregates in the second population. The aggregates in the second population evolve de/aggregation acts that, in the new form, break as many of the control policies as possible. This should lead to control policies that are increasingly resistant to changes in the machine aggregates. Second, we will incorporate evolution-of-development mechanisms to soften the disruption in the task is minimal; later in evolution, aggregation will occur increasingly early during the task, providing a developmental gradient during which behavioral discontinuities can be ironed out. Third, we will enable the separate, specialized neural network controllers of the individual machines to likewise gradually fuse into a single, more general controller in the aggregate. Finally, we will investigate one possible application of this work: a completely soft living computer, in which logic gates are realized as colliding, attaching and detaching blastula.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Sep 01, 2023
- Source ID
- W911NF2310327
Entities
People
- Josh Bongard
Organizations
- Army Contracting Command
- United States Army
- University of Vermont