Biological Robots: Perspectives on an Emerging Interdisciplinary Field
This does not mean that they are useless. Biological robots, like many of the soft robots reported in this journal, challenge conventional views of robotics. They force us to think creatively about how to achieve even the most basic behaviors. The knowledge we gain from our attempts to optimize their design, formalize their control, broaden their functionality, and enhance their intelligence, will eventually become embodied in future useful technologies, whether they are composed of cells, steel, or silicone.
Like every other endeavor, science is not exempt from the limits of human cognition. Primary among these limits is dichotomous thinking: Making sense of nature, and channeling what we know into embodied machines or disembodied neural networks, is easier if we impose distinctions on natural phenomena: for example, organisms have separable bodies and brains, and sense organs and actuation systems, and brain regions, and so on. However, nature need not, and does not respect human attempts to draw boundaries. What follows are some of the most obvious distinctions the scientific community has attempted to impose on nature, and how technological and biological advances are increasingly corroding them.
In genetics, a clear separation between genotype and phenotype has been imposed ever since genes were first hypothesized and then discovered. However, advances in epigenetics and synthetic biology are increasingly revealing that genes, environment, and phenotype are often more usefully thought of as a coupled dynamical system. Xenobots, for instance, support the idea that a developing organism represents a trajectory through an attractor space. The “default phenotype” usually arrived at in response to environmental signals experienced by the organism in its natural environment is but one attractor in this space. Environmental change, ectopic perturbation, or even AI-designed tissue rearrangements can push a collection of cells into stable adult forms completely different from the one usually observed in nature.
In robotics and AI, Cartesian dualism—the West's most famous exemplar of dichotomous thinking—has biased the kinds of scientific questions we ask about intelligence, and the kinds of engineering approaches we take in attempting to create artificial intelligence. Indeed, the very bicameral nature of the robotics and AI communities illustrates how Cartesian dualism warps the field(s). Evidence is now mounting that Cartesian dualism has driven AI into a local optimum: state-of-the-art non-embodied AI is just as vulnerable to adversarial attacks or out-of-distribution environments as it always was,60 and calls to incorporate causal reasoning into AI are growing61 because non-embodied deep networks cannot cause effects and reason about the results.
Our self-replicating xenobots upend a third form of dichotomous thinking that is poisoning computational research: the distinction between tape and machine. This distinction goes all the way back to Turing's original formulation of a theoretical nonhuman “computer.” However, there is no clear “tape” and “machine” in our self-replicating xenobots: nowhere is there some formal instruction to “find loose cells and push them into copies of yourself.” Indeed the xenobots' ability to self-replicate emerges from the collapse of all three forms of dichotomous thinking mentioned above. First, the form and function of the xenobots arise from complex feedback loops between their “genotype” (genetically unmodified frog DNA), phenotype (shape and movement pattern), interoceptive environment (cells and their neighbors within a single xenobot), and exteroceptive environment (the replicative raw material of dissociated cells, and other xenobots).
Second, the “body” of a xenobot is not a binary property during self-replication: there are simply smaller and larger piles of frog cells, and less- or more-motile piles. And, a xenobot's “brain” is the net result of aneural intercellular electric, mechanical, and chemical communication. (Thus, also providing another challenge to Cartesian brain/body distinction by reminding us that there is a diversity of intercellular architectures that can collectively drive increasingly complex behavior.) Third, the geometry of each xenobot dictates how it moves and how, or whether, it contributes to replication: in other words, “the shape is the tape.”
Xenobots are thus an ideal guide for leading us, step by step, into the increasingly deep waters of total morphospace. We must leave behind the shallows, where animals and robots with distinct bodies and brains, genotypes and phenotypes, tapes and machines, swim. Such agents do not actually exist. They only exist in our imaginations because they are the easiest kinds of agents for us to understand. We must leave them and dichotomous thinking behind, and instead learn to swim in the deeps, where real animals reside, and where really intelligent machines will reside. Such animals and machines are ever-shifting patterns of intricately interdependent and fractally arranged bodies and brains, formal descriptors, and physical structures, made that way by the ever shifting currents of natural and artificial selection: it is admixture, all the way down. Such creatures may yield to our understanding, but they will not yield to our attempts to divide them.
Developmental biology began as a study of the phenomenology of specific embryonic model systems. Modern developmental biology, however, comprises not only embryogenesis but also regeneration in adult organisms, metamorphosis, and increasingly—digital and synthetic morphogenetic systems.62–65 What binds these diverse biological examples together into a field is the ability to go beyond specific instances toward questions of multiscale control: How do single cells (which used to be organisms themselves) cooperate toward invariant form and function?66 How does this process increase complexity while resisting external perturbations and harnessing noise, and how does evolution give rise to self-assembling complex systems with both built-in and learned behaviors?
Exclusive focus on the N = 1 example of life provided by Earth's specific phylogenetic tree, with its baggage of frozen accidents, obscures deep principles of life-as-it-can-be.67 Testing our theories with data outside the data set of forms that generated them is an essential aspect of any science, crucial not only for possible exobiology but also to gain a deeper understanding of evolvability, plasticity, and the relationship between genomically specified cellular hardware and the physiological software of life. Understanding the rules of life (moving from Earthly zoology/botany to a fuller option space of possible living things) requires us to recreate and analyze novel living constructs never-before existing on Earth.
Thus, the study of bioengineered forms and their multiscale control policies are an important part of a maturing developmental biology. Xenobots are a good example in which to examine the converse blurring of lines between classical robots and organisms. They challenge us to ask key questions about what we mean by “robot” and “machine,” and whether those binary categories really facilitate understanding and progress.68,69 They challenge us to expand traditional notions of a “program” beyond linear code written by a human to probabilistic, parallel, naturally evolving control policies embodied in biological components that enable the highly flexible, adaptive interoperability of life seen in chimeras and biohybrid constructs.70
Cells are routinely thought of as molecular machines by cell biologists and bioengineers tasked with modifying their behavior;71,72 one of the key challenges of the coming decades is to develop frameworks (or borrow them from information science, cybernetics, computer science, and behavior science) with which to understand the principles by which these remarkable, plastic, evolved machines robustly scale to solve morphogenetic and physiological problems. Developmental and regenerative plasticity73 clearly reveal that multicellular collectives can handle large degrees of novelty; how does evolution result in robust plasticity, and how can we take advantage of these architectures for improved robotics and artificial intelligence? Xenobots reveal new ways to think about control, design, and the multiscale behavior across the spectrum of natural and artificial systems.74–76
Are xenobots engineered or natural products of what Xenopus cells do already? Yes, both. Engineering is not just about adding new ingredients, such as DNA plasmids for synthetic biology circuits or nanomaterials (although these will certainly be added in the future). Xenobots reveal a new strategy for the bioengineer's toolbox: programming desired form and function by releasing natural constraints. In xenobots, nothing was added; instead, cells were liberated from the normal constraints of the rest of the embryo. This led to the discovery that cells' baseline capability is not merely to form a boundary that keeps pathogens out of the organism. Instead, this mundane two-dimensional lifestyle is forced on them by the instructive interactions of other cell types. When allowed to express their own multicellularity, these epithelial cells do not die, form a monolayer, or wander aimlessly. Instead, their baseline behavior is to form a self-motile proto-organism with individual and collective behaviors, such as kinematic self-replication.9
These types of form and function were not obvious from any first principles or from their wild-type genotype and illustrate the need to understand what it is that the evolutionary process taught the X. laevis genome;77,78 it was not only how to make a frog or a tadpole. There has never been selection to form a functional xenobot, which raises the important question of what defines the class of constructs that cells can make, beyond the default configuration supported by eons of selection forces. What else do these, and all the other, cells know how to do, in novel circumstances? Of course, their capabilities are explainable, after their discovery, as genetically specified features (such as cilia and adhesion proteins) working in novel ways through the laws of physics.
We encourage readers to make (and register in advance) predictions of other capabilities that cells such as these will exhibit—the degree of their plasticity and the specific forms and behaviors of which they might be capable. The predictions of their capacities and limitations (so-called developmental constraints) will be interesting to test against the rapidly emerging data in this field. The fact that these are no more predictable ab initio from genomic information and current models than the primary shapes of animals and plants is an important challenge to the field.
The xenobots' ability to assemble the next generation from cellular material in their environment sheds an important light on the notion of control, both in the evolutionary and in the engineering context. Indeed, both use the same strategy: reliance on the competency of their components. Engineers make the first generation of xenobots by relying on the willingness of cells to get together despite novel circumstances and create a motile coherent agent. These xenobots create their next generation by exactly the same process: they assemble the cells and impact the size and composition of the collective, but then take advantage of the competency of cells to do the rest—guided self-assembly (i.e., behavior shaping), not micromanagement is the way that we, and the xenobots themselves, make more xenobots. Indeed, evolution itself relies on the competency of cells, tissues, and organs to solve problems.79
Making a copy of a biobot does not have to be micromanaged any more than the original biobot needed to be three-dimensional (3D) printed. This does not resemble traditional robotics because the history of engineering has relied on passive materials. Progress is revealing a continuum of control strategies appropriate to active, computational, and agential materials (Fig. 3): building things with parts that have context-sensitive activity, computational capacity, and behavioral agendas that optimize various parameters in local problem spaces. Evolution likewise relies on its components to provide more than reliable passive form (LEGOs that hold their shape but do nothing else)—instead, it largely shapes signals and biophysically implemented incentives to modulate the behavior of cells, tissues, and organs.
Biologists, roboticists, and intellectual property lawyers will have to become comfortable with workflows in which we are literally collaborators with our artifacts, parts of which will actively build the next level of the construct in ways that are not fully captured by a craftsman's action protocol. This has massive advantages, for rational design and for evolution, because it allows both to work in a simpler space of incentives and inputs, not microstates. It also provides extensive challenges because our science of prediction and control of systems with a multiscale competency architecture is still nascent.
The existing variants of xenobots have not yet been provided with edited genomes, new materials, and so on. Counter to established intuitions, the kind of control expected of robotic platforms (machines) does not require genomic editing. Computer science achieved the information technology revolution by moving from programming by hardware rewiring to software control via inputs. Evolution discovered this trick via physiological software very early on,85 and recent advances in the control of growth and form show how much anatomical change can be implemented by interventions that do not change the underlying hardware (genetics).86
Induction of appendage regeneration, production of structures belonging to other species despite a wild-type genome, induction or normalization of cancer, and so on, can all be induced by biophysical information signals and often invisible to canonical molecular biology tools.87 A key aspect of the early generations of xenobots is precisely that they reveal the plasticity of cells' ability to explore morphospace and behavioral space with the same genome. This is an expansion of “epigenetics,” toward a better understanding of plasticity and the actual products of evolution.
The practical applications of such biobots are numerous. Major advantages are their biodegradable, soft nature, and the built-in competencies of cells for sensing, metabolic/biochemical activity, and actuation in ways that are far beyond today's engineering efforts. Most obvious are the useful living machines, which could microsculpt bioengineered tissues in vitro, perform sensing or cleanup in many industrial processes, environmental cleanup, or even in-body biomedical applications with respect to injury, infectious microbes, and cancer. They also provide an inexpensive safe model system for education in STEM efforts (such as Frugal Science).
Deeper impacts include the use of this platform as a sandbox in which to crack the morphogenetic code and move closer to the endgame of an “anatomical compiler.” Work in biobots will help address the gaps in our ability to predict and control complex anatomical shapes66,88 for applications in developing signaling protocols to induce cells in the body to achieve correct regenerative morphogenesis. An essential aspect of such strategies to address birth defects, traumatic injury, and errors of multicellularity such as cancer and aging is learning to control the shape toward which cellular collectives cooperate. Biobots are an enabling platform to achieve Feynman's dictum—to truly understand morphogenetic systems by building them ourselves.
Benefits will also accrue to computer science and robotics in inorganic media because the fascinating features of biobots do not derive from any magic of protoplasm. Rather, they derive from their multiscale competency architecture and other principles to be discovered. Current robotics is safe from cancer (defections of components from system-level goals) but is brittle and lacking general intelligence. The opportunity is to borrow from nature, not the specific genes and pathways comprising today's developmental biology textbooks, but the ways in which non-neural systems solve problems in diverse problem spaces by scaling their basal homeostatic properties into larger goals.89,90 The implications for AI and behavioral science in general concern biobots' degrees of proto-cognitive sophistication. No claims have yet been made for the degree of agency they can exhibit (this characterization is currently under way), but it is clear that humans' intuitions about recognizing intelligent behavior are highly limited by the training set of familiar animals behaving in 3D space.79,91
The many surprises in the basal cognition literature92 reveal that we are fundamentally not good at recognizing possible intelligence in navigating physiological, morphological, and other non-obvious spaces; a rich research program exists around efforts to understand xenobots' unconventional intelligence (at the very least, as a warm-up to exobiological and AI advances). Finally, the long-term implications of this research program address a key component of avoiding existential risk for humanity: better learning to understand and predict the behaviors of systems composed of competent components (emergent properties of collective intelligences).
This technology also has important implications for ethics.93,94 These are not limited to familiar biosafety concerns, as these technologies are much easier to contain and have much better safety profiles than existing efforts in synthetic bacteria, engineered viruses, and genetically modified crops and other organisms. Xenobots, for example, live only in very specific environments and are highly dependent on humans for their manufacture; they biodegrade rapidly—unlike many of today's synthetic biology efforts that replicate readily in the biosphere. An aspect often forgotten is that we cannot just evaluate potential risk as if the status quo was excellent, and our goal was merely not to make things worse. Human and animal suffering worldwide is enormous and due to a poor understanding of system-level biological control.
There is a massive opportunity cost of not pursuing this research because of its potential to advance regenerative medicine and improve quality of life worldwide. Any risk calculations of such technologies must balance against the moral duty to pursue scientific programs that can alleviate unmet environmental and biomedical needs and health care disparities. We ethically owe it to victims of birth defects, injury, cancer, and degenerative disease, and so on, and to the animal model systems currently used for biomedical research (which could be partially replaced by simplified constructs), to better understand how to influence cells to achieve complex structural and functional states.95,96 Another important aspect of ethics in this kind of work is that it forces us to confront rules based superficial criteria—past distinctions of origin and composition which at one point might have been a reasonable guide to how one ought to treat a given system but are not fundamental and based only on past limitations of technology and imagination.
Human society has a long history of moral failures toward numerous types of human and nonhuman animals based on distinctions that are now widely recognized to be irrelevant. By forcing us to search for deeper essential meanings of categories that drive moral concern and responsibility (agency, intelligence, etc.) in unfamiliar novel guises, we will be better placed to more appropriately implement compassion in our relationship to a truly diverse community of beings.
Robotics is an attempt to engineer useful, partly or fully autonomous artifacts that remind us in some way of organisms. Increasingly, this research is breaking with the classical conception of a robot as necessarily made of passive inorganic components and made or designed piece-by-piece by humans at every step. We believe that the essential nature of robotics is not limited to a specific kind of material, origin story, or type of control. However, this stance necessitates a re-evaluation of terminology as multidisciplinary teams bring field-specific techniques to bear on new questions. As we lay out, this terminology remains in a state of flux, as audiences continue to view these questions through their own unique lens.
Recent work in biohybrid constructs—machines made of cells3,34,35,75,97–107—is giving rise to a very dynamic emerging field. We do not here comprehensively review all the existing work (see Ebrahimkhani and Levin,65 Kamm and Bashir,72 and Kamm et al108), but rather to provide our individual perspectives on specific issues central to a number of communities in the biological and engineering sciences. Nor are we attempting to review the growing literature on robot design automation,58,109,110 which has clearly shaped our own contributions and perspectives and is constantly evolving.
We suggest that many new research programs challenge the traditional lines drawn between a machine, a robot, and an organism. Living, “active,” “smart,” or “agential” materials (depending on the author) have much to offer robots and engineering disciplines. Simultaneously, materials science research can provide insights into basic developmental biology and bioengineering as cells and tissues assemble during the design process. From the opposite end, simulation and AI-driven design have the potential to enhance traditional biology programs, as sim-to-real approaches continue to accelerate in vivo and in vitro discovery.
Within the group, the authors vary in specific terminology and system label usage (e.g., robot vs. organism vs. xenobot), whether they characterize designs as a group of skin cells or something new, and when speculating about the degree of sensory motor communication possible in the system. These issues remain areas of active investigation. Finally, we understand where disagreements arise among fields and hope the views presented here help frame the research program to the larger scientific community.