Toward Growing Robots: A Historical Evolution from Cellular to Plant-Inspired Robotics

This paper provides the very first definition of “growing robots”: a category of robots that imitates biological growth by the incremental addition of material. Although this nomenclature is quite new, the concept of morphological evolution, which is behind growth, has been extensively addressed in engineering and robotics. In fact, the idea of reproducing processes that belong to living systems has always attracted scientists and engineers. The creation of systems that adapt reliably and effectively to the environment with their morphology and control would be beneficial for many different applications, including terrestrial and space exploration or the monitoring of disasters or dangerous environments. Different approaches have been proposed over the years for solving the morphological adaptation of artificial systems, e.g., self-assembly, self-reconfigurability, evolution of virtual creatures, plant inspiration. This work reviews the main milestones in relation to growing robots, starting from the original concept of a self-replicating automaton to the achievements obtained by plant inspiration, which provided an alternative solution to the challenges of creating robots with self-building capabilities. A selection of robots representative of growth functioning is also discussed, grouped by the natural element used as model: molecule, cell, or organism growth-inspired robots. Finally, the historical evolution of growing robots is outlined together with a discussion of the future challenges toward solutions that more faithfully can represent biological growth.


iNTRODUCTiON
The generation of a biological organism involves three tightly connected processes: growth (changes in mass), remodeling (changes in material properties), and morphogenesis (evolution of the shape) (Taber, 1995). Growth can be interpreted either as a physical process in which there is a permanent increase in mass, typically with body enlargement or elongation, or it can be seen as an abstract process that describes the increase in system complexity (e.g., by the enhancement of knowledge).
Growth involves the cellular activity of both animals and plants; however, the evolution of organ isms in these two kingdoms is completely different. Animals have multiple stages of evolution, they grow until maturity, while plants grow indefinitely, mostly for their entire life. Growth in plants is typically restricted to localized areas, called meristematic zones of roots and shoots, whereas animals' growth zones are distributed all over the body (Soni, 2010). Animal growth is also known as "determi nate" growth, since trajectory and asymptotic size are usually genetically defined and environmental influence has a limited impact. Natural selection through evolution has driven the final size and stage of maturity to better satisfy the mechanical constraints imposed by the ecological role of the animal (Sebens, 1987). modify their mass by adding new modules to a main entity, while selfreconfiguration can be considered as an extra feature, not strictly related to growth. In fact, not all reconfigurable robots can be considered as growing robots since reconfiguration can also happen in a closed system without an increase in volume or mass of the robot.
Modular robotics is the main popular approach used by engineers to translate the morphological evolution of biological organisms into artificial systems. A more extensive understand ing of biological processes involved in growth mechanisms can help in improving bioinspired approaches to generate artifacts closer to biological growth.
Bioinspiration has already helped to find solutions for the adaptation of robotic bodies and behavior to environmental changes (Pfeifer et al., 2007). The study of neural systems and evolutionary processes has guided heuristic solutions to problems that are too complex for analytical methods and toward the "intellectual" growth of machines (Goldberg and Holland, 1988;Floreano and Mattiussi, 2008). Because of their particular features and difference from animals, plants have already contributed to robotics through, for example, sensing (Lucarotti et al., 2015), materials (Li and Wang, 2016), actua tion (Sinibaldi et al., 2014), and control strategies . Such perspectives have given rise to plantinspired robotics.
This new field of robotics is attracting interest by roboticists and engineers Wahby et al., 2016) and has also brought the concept of plant inspiration to modular robotics. For instance, Soorati and Hamann (2016) mimic plant growth following the traditional approach of modular robots by arranging a set of Kilobots (Rubenstein and Nagpal, 2010) in a branchlike shape according to light distribution.
The utility of a robot that evolves and grows lies in its enor mous potential adaptability. In unstructured environments, where constraints are not well known beforehand and where access is limited or not recommended for humans (e.g., after an earthquake), or to locate objects that are difficult to reach, a robotic system capable of quickly adapting its locomotion, mor phology, and functionality according to the environment and the task can be much more effective in terms of intervention time and working efficiency.
This paper reviews the milestones in the concept of growth in robotics. Section "Toward Growing Robots" focuses on the historical approaches and on the recent results achieved in the area of plantinspired robotics, specifically on the first robot inspired by roots that move in soil by growing its body. In Section "Bioinspiration Toward Robots that Grow, " we discuss a selection of robotic solutions grouped according to the biological system model and critically analyze the concept of growth implemented. The classification adopted in this paper to categorize growing robots is based on the complexity, or hierarchy, of the natural element addressed: molecules, which are the elementary units for life; cells, which assemble forming tissues; and an entire organism which is composed of the elementary units. Figure 1 gives an overview of this classification. Finally, in Section "Discussion and Conclusion, " we provide a general discussion on the future challenges.

TOwARD GROwiNG ROBOTS
The concept of growing machines can be attributed to John Von Neumann, when back in the mid twentieth century he discussed the concepts of selfreproductive automata and "complication. " The "concept of complication" expresses the idea that natural organisms reproduce themselves without decreasing complexity, but instead, through evolutionary processes, new systems are more complex than the parent systems. He speculated if and how such a concept could have been reproduced into automata. Starting from Turing's theory of computing automata and Ulam's suggestion about a cellular model (Beyer et al., 1985), Von Neumann (1951) formalized his idea on how an automaton can generate another identical automaton starting from a reservoir of floating elements.
Considering the high level of complexity of natural systems, Von Neumann suggests an approach based on the translation of natural processes into artificial systems by two steps: (I) break ing down the problem into subproblems, e.g., a single organism is made up of many elementary units; (II) understanding how these individual elements are organized and contribute to the functioning of the whole system. This idea of decomposing a complex organism into smaller and simpler units was later adopted for the first time in robotics by , who presented his concept of a dynami cally reconfigurable robotic system (DRRS) implemented in CEBOT (a cell structured robot). This seminal work introduced the concept of a robot not as a single unit but as a distributed robotic system made up of separate autonomous and heterogene ous units, called cells. These robotic cells can be functionalized and are able to communicate with each other, to approach, connect, and disconnect automatically. The composition of cells forms a single module, like a tissue; several modules are assembled to make a single structured robot. The system can adapt its structure to the environment and perform a task by a different combination of modules. By adopting the same strategy of recombination, such a system shows selfrepairing and fault tolerance abilities and is able to continue operating even when a failure occurs. The theory at the basis of these decomposed self assembling robots went far beyond the real implementationsseveral versions of CEBOT were implemented (Mark I, II, and III)-which, however, were made up of only two modules . Despite this, CEBOT was the first prototype of a modular robot, opening up the area of "cellular robotics, " also named "modular robotics. " Cellular robotics is a broad research area where simple or primitive intelligence (e.g., social insects) are studied and adopted to organize a group or community of robots in terms of their coordination and cooperation (Fukuda and Ueyama, 1994). The main characteristics include: modularity, representing cellular structures; simple functionality, each cell has limited capabilities; and decentralized control, the intelligence of the robot emerges from the interaction among cells. Cellular modules are mainly prefabricated mechatronic cells whose assembly and combination constitutes a single individual (a cellular robot). A community of individuals who interact with each other constitutes a cellular robotic system. Based on this concept, cellular robotics also incorporates the field of "swarm robotics" (Beni and Wang, 1993;Fukuda and Ueyama, 1994).
Modular robotics has made a step toward Von Neumann's idea of selfreproductive automata possible, through the exploration in robotics of selforganization and selfconstruction properties adopted by living systems. After Fukuda's DRRS, the challenge was to obtain a system not limited by the initial design, but which could extend itself through construction methods, and the combination or connection of cells. The dynamic reconfiguration enables the structure, size, and functions of the robot to be adapted according to the task. Selfassembly, selfreconfigurability, and distributed control became the features of many works carried out by the robotic community. Which of these works actually incorporate properties of growth is discussed below.
An alternative approach to cellular robotics was proposed by Lipson and Pollack (2000) with the concept of a continuously selfdesigning reconfigurable robot. They proposed to combine the power of evolutionary computation to design the body and develop the control at the same time, with additive manufactur ing to fabricate it. On the basis of their approach, a single complex robot is composed of three main units: one unit computes the design, the second unit fabricates the body, and the third one is the result of the process, i.e., the body produced. The idea is to produce a robot that without human intervention is able to self produce control and design of the body that best satisfy a task.
Lipson and Pollack also aimed to develop a robot made of recycled material and components that could be fused and readapted for new morphologies. They implemented an initial idea based on an evolutionary process capable of producing a combination of linear actuators and bars connected with free joints for the body morphology which can travel for a certain distance. In this process, a neural network is evolved for the control and the outcome of the design is then produced through a commercial 3D printer machine. The components need then to be assembled by hand, motors are added, and then programmed. The main limitation of their approach was the technology avail able. Additive manufacturing equipment seemed impractical as part of a robot and assistance by humans was essential to assemble the components, although the design of the robot was automatically generated by a computation unit. On the other hand, in Lipson's vision, additive manufacturing can be seen not only as an alternative to classic manufacturing techniques (e.g., molding or assembling), for instance, in the creation of the body of a robot [e.g., Bartlett et al. (2015)], or to assist and enhance robot functionalities [e.g., used to build tools or grippers by the robot itself (Wang et al., 2014a)], but also as an integrated and essential component of the robotic system. It enables unpredictable structures to be created, defined by the control, and to explore designs that would otherwise be poten tially difficult to manufacture with classical approaches.
The first methodology to evolve body and control together was proposed by Sims (1994), where a series of different virtual creatures compete in simulated threedimensional worlds. Evolutionary selection rewards those with the highest scores, letting them survive and reproduce. Morphology and behavior evolve simultaneously. The genotype is represented by directed graphs describing the instructions for the creatures' growth. A similar approach (treebased genotype) was later used by Marbach and Ijspeert (2004) in their Adam simulator which coevolved configuration and a PD control for homogenous chaintype modular robots. Their objective was to develop a simulator to define the configuration of physically implementable modular robots with a fixed and known design, together with the control for the appropriate locomotion of the defined structure. von Haller et al. (2005) coevolved the body configuration of an underwater modular robot with a neural network controller based on a central pattern generator.
Another perspective of the evolution of virtual creatures is to provide an abstraction from a real robot (Bongard, 2013). This approach converged into "evolutionary robotics. " The main goal of this methodology is to reproduce evolutionary processes to find the optimal robotic structures for a specific task that do not necessarily resemble existing creatures. This field of robotics also exploits the "evolution of development, " which means the process by which a creature evolves from the embryonic stage to maturity, while being affected by evolution over time (Bongard, 2013). Bongard (2011) proposed the evolution of an anguilliform virtual creature that mutates into a legged robot in the early stage of evolution. Locomotion gaits evolved with morphological changes, and he observed that creatures evolving from anguil liform to legged creatures performed the transition more rapidly with gaits that were also more robust compared to creatures that already had legged bodies.
Evolutionary robotics also represents a biological instrument for understanding morphogenetic properties and subsequently the evolutionary transitions of real creatures. Evolutionary robotics has thus been included in the broader field of "morpho genetic engineering" (Doursat et al., 2012), which is an evolution of cellular robotics since it is based on a deeper understanding of how organisms and populations reliably accomplish morpho genetic tasks. It also examines how selforganization capabilities can be translated into engineered systems and which principles can achieve a morphogenetic system. The basic principles are still the same as cellular robotics: modularity and decentralized control; although it is endowed with advanced tools and algorithmic approaches, such as those offered by evolutionary robotics. These approaches simultaneously explore the design and control of autonomous systems, which are capable of devel oping complex and heterogeneous morphologies and functions in a decentralized manner.
The results in this field have demonstrated the feasibility of fabricating an evolved modular creature (Hiller and Lipson, 2012) and the transferability of an evolved control onto a physical not evolving robot (Koos et al., 2013), putting significant effort in the computation of the robot design and morphological evolution in simulation. The approach adopted by evolutionary robotics has also been used for the evolution of virtual soft bodies (Joachimczak and Wróbel, 2012), especially for the investigation of material properties and their effects on robot behavior (Rieffel et al., 2014;Corucci et al., 2016). All of these works, and many others, demonstrate the feasibility of obtaining an intellectually growing and adaptable robot. However, the step from simulation to the physical creation of a growing-not only intellectually but also morphologically-completely autonomous working robot is still ongoing. Again, the main challenge seems to lie in the physi cal transduction of evolutionary processes for the morphological evolution and growth of a real robot.
A step further toward the physical implementation of robots that grow has been achieved by taking inspiration from the concept of growth in plants (see OrganismInspired Robots). The starting approach originated from the observation of the high adaptability of plants to harsh environments and their ability to colonize soil, skills that can be translated into artificial devices for soil penetration and exploration (Mazzolai et al., 2008). By investigating different strategies adopted by plants, a series of energetically efficient technological solutions to movement in soil have been proposed. For instance, in Sadeghi et al. (2013), the release of dead cells from the root tip was implemented with a mechanism of skin eversion from the inside of a tubular shaft to the outside. This was done by creating an interface between the shaft and the soil and demonstrated the reduction in friction perception in soil penetration. In Del Dottore et al. (2016), the oscillatory movement that is also actuated at the tip level in plant roots was used in a robotic root to demonstrate the reduction in forces needed to penetrate into soil with respect to a straight penetration. The particular features of roots and shoots that enable them to grow from their apical regions by cell division and elongation were first translated into a rootlike device that deposits a filament in a circular manner, preserving the contact of the filament with the artificial tip, thus providing sufficient force for pushing the probe into soil (Sadeghi et al., 2014). Sadeghi's study paved the way for a new generation of robots inspired by plants which also provides an alternative to cellular robotics aimed at physically implementing the concept of the "selfbuilding" robot.

BiOiNSPiRATiON TOwARD ROBOTS THAT GROw
This section provides an overview of the robots that have been inspired by the mechanisms of growth implemented by biological systems and in which a bonding strategy has led to this feature being physically implemented. The main categories are modular robots, with physical selfassembling capabilities, and plant inspired robotics where robots that grow imitate the strategy adopted by plants of growing from the tip.
We have categorized the selected robots according to an increasing hierarchical dimension of natural elements, starting from molecules up to a single complete organism (see Figure 1; Table 1).

Molecule-inspired Robots
This category includes robots inspired by molecule polymeriza tion in DNA formation (Figure 2). This process is the basis of all evolutionary processes, including growth. The artificial implementations are based on stochastic collisions of particles-Brownian motion (Einstein, 1956)-which are not active when disconnected but acquire functionalization once a bond is estab lished, leading to the ability to assemble, seen here as growth.
One of the first examples comes from Griffith who fabricated small entities capable of selfassembly (Griffith, 2004). His goal was to reproduce the assembling capabilities of biological sys tems at the nano/microscale in order to devise new fabrication technologies that would enable selfassembly threedimensional structures to be created that would perform errorprevention, errorcorrection, logic, selfreplication, and selfrepair by the replication of part assemblies. The result of his work was a self replicating system composed of a finitestate machine randomly moving on a floating table which could physically connect or disconnect with/from each other in response to neighboring communications (Griffith et al., 2005). In this implementation, each part does not need to store a copy of the entire structure but queries neighbors in order to test selfsimilarity and its position in the growing replicant. These machines were able to compose strings in 2D-sequence of variable numbers of modules only in one line-and to create copies of these strings. The growing entity is a single string, which starts from a single module and increases in size and functionality thanks to the assembly of other modules.
In White et al. (2004), an approach based on passive modules moving on a floating table that bind stochastically was proposed. White proposed two geometries, a square and a triangle, for 2D motion and assembly. Binding starts from a single active module (a seed) which, according to internal rules, can activate or deactivate bonding sites. The motion is based on statistical mechanics and on attraction properties of active bonding sites. Once a bond is established, the resulting structure is then able to activate global and distributed sensing, actuation, and computa tion. Bonding sites are implemented with electromagnets and the structure starts to grow from a seed module. Two 3D versions of the same principle have also been implemented (White et al., 2005): one implementation uses magnets for bonding, as in the 2D case; the second implementation instead uses the activation or deactivation of internal pumps to form or detach bonds. In both 3D implementations, modules passively move in a liquid which through agitation introduces randomness into the system. A module fixed at the bottom of the tank acts as a seed and attracts the floating modules.
The same floating principle was proposed in Haghighat et al. (2015Haghighat et al. ( , 2016, where Lily robots assemble themselves following graph grammar rules, which specify conditions of bonding when randomly mating according to a target structure. These robots can be configured into different shapes. The Programmable Parts proposed in Bishop et al. (2005) are triangular robots that passively float on an air table and bind to each other upon random collisions. Chemical reactions are imi tated by graph grammar rules employed to establish or detach a bond. However, the stochasticity and assembling capability are insufficient to make it a growing robot. In fact, it is only able to assemble itself in a single form which is strictly dependent on the geometry of the module. The combination of a different morphology with the stochasticity of the environment has been demonstrated to affect the clusterization of modules enormously (Miyashita et al., 2011). In addition, the simple effect of morphol ogy has been analyzed, for instance, in Miyashita et al. (2009) or Nakajima et al. (2012), also demonstrating in these cases the different clusterization according to different geometries.
Moleculeinspired robots can artificially implement a natural process without performing complex tasks, although they are extremely dependent on the module geometry that defines the assembly and on the floating environment that provides the energy for locomotion. They are thus mainly limited to PoCs.

Cell-inspired Robots
This category potentially includes all modular robots, since they derive from the idea of replicating cell structures. Primarily, ideologically they resemble cells; however, they do not imple ment the biological functionality of growth by cell division and selfproduction.
There are several reviews on modular robots with self assembling and selfreconfigurable capabilities (Gilpin and Rus, 2010) (Groß and Dorigo, 2008) (Yim et al., 2009). Modular robots are usually classified depending on the architecture (lattice, chain, mobile, and hybrid) or method of reconfiguration (deterministic, stochastic). Our interest is in the growth functionality that these robots implement. Our aim is not to provide another review on modular robotics but to select only those modular robots that show growing capabilities by means of selfassembly, which for instance purely lattice or chain modular robots do not have in general (Figure 3).
However, in terms of modular robotics, the focus has been on defining strategies for system reconfiguration, which uses the concept of swarm intelligence to provide more independence to every single module and to create strategies for collaboration among modules. In addition, fabrication technologies have been aimed at achieving mobile autonomous modules that physically connect with each other. Such systems, which mostly move and connect forming 2D configurations [e.g., SwarmBot (O'Grady et al., 2005), ULGEN (Ercan and Boyraz, 2016), XCELLs robot (Hong et al., 2011)], are beyond the scope of the present review. This is because they mainly simulate the collective behavior of colonies of individuals, where a single individual already has an enhanced level of autonomy and the assembly does not provide a real augmentation of different capabilities. As mentioned in Section "Toward Growing Robots, " these works in swarm robotics resemble cellular robotic systems more than cellular robots. CEBOT , which was built for manipulation purposes in industrial environments, is thus the precursor of modularity and distributed control (see Toward Growing Robots) although it did not achieve full functionality. CONRO (Rubenstein et al., 2004) shows its growing capability by the selfassembly of two segments, composed of two modules each. Segment docking needs to be performed with the appropri ate alignment of malefemale mechanisms and the detection of the two segments is constrained by IR interface alignment. The PolyBot (Yim et al., 2002) assembly strategy was designed very similarly to CONRO, which in addition to CONRO shows the docking of a sixmodule arm to a single module positioned at a fixed and known location.
Its more advanced successor CKbot (Yim et al., 2007) uses vis ual feedback to locate neighboring disconnected modules. Each disconnected cluster of modules needs a camera module to guide its docking, and all clusters search for the others. Visual feedback had already been used by MTRAN (Murata et al., 2006), where a single camera module is positioned on a stationary cluster to calculate the position and orientation of the disconnected cluster of modules and guides them toward the docking position.
SMORES (Davey et al., 2012) is a hybrid system which can connect two modules by magnets closely positioned to autono mously induce the connection (search and approach are not described). MBlocks (Romanishin et al., 2013) are independent robotic blocks which assemble by magnetic edges and faces, using a unidirectional reaction wheel (flywheel) to create an approximate impulse of torque. When a block is not aligned with a goal position, the module actuator will use high torque leading to a random movement; instead, if the module is aligned with the goal position, torque is used, leading to a controlled rotation toward the goal. Qiao et al. (2014) show how a mobile module docks itself to a module fixed at a defined position. Wei et al. (2011) present another good example of cell inspired growing robot (Sambot), although it probably lies on the boundary between swarm robot and a selfassembling mobile robot. Through selfassembly, this multirobot system can form a variety of robotic structures with locomotion capabilities. Docking is achieved with IR interfaces and mechanical hooks. Mechanical hooks are also used by Roombot modules (Spröwitz et al., 2014) to connect to each other. Spröwitz shows the ability of the two modules to connect in order to pass a convex edge; however, modules need a precomputed set of steps and also need to be connected to a grid with passive connectors embed ded in the floor.
In cellular robotics, most of the focus has been on docking strategies, for searching, approaching, and connecting discon nected modules. To obtain a bonding mechanism, a module needs to apply a sufficient force to connect to another one and hopefully lift it. At the same time, a fast release is required. In addition, strategies for finding the optimal reconfiguration are important. Generally, search and step planning algorithms are implemented on a separate working station and only then are the steps to follow are provided to each module.
Growth, or in this case selfassembly functionality, is not usually a key objective in modular robotics. One example where growth acquires more relevance is the Proteo robot (Bojinov et al., 2000). Proteo is a homogeneous multiagent self reconfigurable robotic system, which according to the authors, is able to grow stable structures. Bojinov configured the robot using local attractors-modules called seeds-which produce a gradient of attraction toward the desired global configuration, like chemical gradients guiding migrating cells (Meinhardt, 1993). However, the concept of growth is used in Proteo as a strategy-the directed evolution of the shape-not for the assembly but for reconfiguration, to change the robot shape from one configuration to another. Moreover, Proteo has shown its capabilities in simulation and not in a physical implementation, and it does not implement a selfassembling or alternatively self building strategy.
All the cases reported in this section have constraints in successfully connecting the modules. In fact, one of the two disjoint components needs to be fixed to a defined position; or both components need to contribute to the assembling process, making it difficult to recognize the actual entity increasing in size, or rather, growing. Moreover, the granularity of growth cannot be tuned: the minimal size is predefined by the module size (discrete process). The assembly is clearly not real growth, but it has been the closest approximation to growth and successful artificial process for several years.

Organism-inspired Robots
This category considers robots inspired by the growth strate gies adopted by an entire organism, either simple or complex, such as animals, fungi, and plants. Unlike cellinspired growing robots, where several modules connect to each other to emulate the assembly of cells, organisminspired robots include single module systems which are able to incrementally add material in specific areas of their body. Sadeghi et al. (2014) proposed the first device capable of verti cally penetrating soil through an inert filament at the tip, thus imitating growth from the tip adopted by plants. The subsequent result of this pioneering work was the creation of a plantrootlike robot able to build its own body structure (Sadeghi et al., 2017) ( Figure 4A). The robotic root hosts a miniature 3D printer inside its tip and uses a thermoplastic filament to build a stable hollow body where layers of fused material adhere to each other and solidify consolidating the structure (FDM-fused deposition modeling approach). System bending is achieved by differential growing, by depositing different amounts of raw material at the opposite sides of the robot. The tip embeds sensors and a control unit with a bioinspired behavior , which drives the direction of growth actuated on the back of the tip allowing the following of attractors or avoidance of repellents. This work represents the first physical solution to selfcreating robots, by integrating additive manufacturing techniques inside the robotic bodies.
The adoption of thermoplastic material has also been proposed by Wang et al. (2014b) who developed a spiderlike robot. In this case, the material is released by the robot for the creation of a dragline. Yet, the continuous filament released by this robot cannot be considered as an integral part of the robot's body since it is an external component used by the same robot for its locomotion-by lateral wheels sliding on the formed  filament-not to morphologically change the body, thus invali dating the function of growth. Hawkes et al. (2017) exploited the idea of body elongation adopted by several biological systems such as fungal hyphae, neurons, or pollen tubes in seed plants. Hawkes et al. imple mented a robotic system that can elongate its body from the tip by skin eversion using pressurized chambers, and by indi vidually controlling pressurization on each chamber, the system can bend for aboveground environment navigation driven by a visual processing controller Figure 4B. In fact, there are fixed pinchlatch mechanisms along the skin, with which higher pres sure unlatch, thus enabling bending on the opposite side.
In the above examples, the main limitation lies in the quantity of material available. This means that the ability to grow is still dependent on and limited by a reservoir of material that needs to be provided to the entity managing deposition or inflation.

DiSCUSSiON AND CONClUSiON
This paper presents the historical evolution of growing robots. We propose a definition of this category where robots physically evolve their artificial bodies by the addition of material. This principle represents a dream for many scientists and engineers, and different approaches have long been proposed in robotics. Ideally, from a longterm perspective, growing robots should be able to selfbuild their own structure and upgrade functionalities on the basis of elements present in the environment, making them completely independent from a reservoir of material or human intervention.
Currently, none of the reported solutions exploit the environ ment, however cellular robotics, with modular robots that are capable of selfassembly (modules are able to autonomously connect to each other) represents the first attempt toward this goal. This research has led to the mechanical improvement of physical connections and to algorithmic approaches for template matching, configuration selection or path planning, to reduce the reconfiguration time and optimize module arrangements for a specific task. The concept of growth has been extended to a more general morphological evolution and is typically treated in these robots with the selfassembly and rearrangement of modules.
Taking inspiration from nature, the elongation of tissue has been imitated by everting artificial skin from the inside to the outside of a tubular body (Sadeghi et al., 2013) or by pressur ized chambers (Hawkes et al., 2017), obtaining rapid growth and bending capabilities. A new approach has also been discussed that led to the first robot beyond modular robotics capable of building its own structure from a supplied thermoplastic material (Sadeghi et al., 2017). The robot is able to manipulate, change material properties, and deposit the material in order to build a structure that can be considered as a channel for communication, which would be particularly useful in rescue applications, for pro viding oxygen, water, to pass through other robots, cameras, or sensors. The tubular built structure also easily provides a passage for the delivery of new raw material and energy necessary for the growing process.
In the examples reported as organisminspired robots, there is a finer granularity of growth compared to the robots in mol ecule and cellinspired robots. Instead of having the minimum growth dictated by the size of a module, organismlike growing robots are able to tune their growth by speeding filament deposi tion (Sadeghi et al., 2017) or chamber inflation (Hawkes et al., 2017) up or down, sparing the need of technically complex and expensive modules in imitating the growth process.
The mechanisms proposed in this category have various pros and cons. For instance, the exploitation of additive manufactur ing in plantgrowthinspired robots enables the exploration of soil and unstructured environments besides the aboveground scenario. On the other hand, an inflating mechanism, for instance, can obtain a faster growth velocity respect to the FDM technique, enabling the robot for a fast reaction in aboveground environment. In fact, in adopting a plotting strategy for material deposition, the process can be quite slow, which can be overcome for instance, by changing the printing technique, e.g., the projec tion of an entire area, or using a material with a different bonding mechanism, e.g., a chemical bond as in hydrogels. However, with respect to fast growth, a slow process can be more beneficial where high impedance-e.g., in soil-or the high risk of damagee.g., unstable structures, or human's tissues-are constraints of the environment where the robot has to move.
Yet, the physical implementation of growing robots is limited by the current difficulties in technologically imitating cell divi sion and absorption of environmental resources.
Selfhealing for instance is a feature embedded in natural tissue enabling the reconstruction of dead or damaged tissue, and is also a desirable feature for growing robots. Selfhealing properties, which are often confused with the more practical concept of "selfrepair" (i.e., maintenance) and implemented in robotics with redundancy and reconfiguration (Murata et al., 2001), have been recently explored in materials (Wool, 2008;Hager et al., 2010;Yang and Urban, 2013). Selfhealing polymers have already been adopted in soft robotics as soft pneumatic actuators-e.g., artificial muscles-and soft hands able to reestablish crosslink bonding through a thermoreversible Diels-Alder reaction after being damaged by sharped objects (Terryn et al., 2017). Also, studies are currently being performed to imitate material micro structures for the dynamic adaptation of the shape, for example using hydrogels (Gladman et al., 2016). Tissue engineering and synthetic biology (Cheng and Lu, 2012) could provide a complementary approach to classic engi neering. In fact, in synthetic biology, the objective is to build new biological systems starting (I) from an existing system and by reducing its complexity (topdown approach)-designing, syn thesizing or recombining DNA to preserve essential and desired functionalities (Gibson et al., 2008(Gibson et al., , 2010)-or (II) building a new system from basic units (bottomup approach) (Schwille, 2011). This approach has led to the construction of in vitro circuits from synthetic DNA switches (Gardner et al., 2000;Kim et al., 2005Kim et al., , 2006, synthetic polymer scaffolds for the development of artificial bones (Song et al., 2004) or actuators powered by biohybrid materials synthesized from natural muscles (Feinberg et al., 2007). This thus demonstrates that synthetic biology offers functional tools for artificial selfassembling and selfreplicating systems. The combination of these tools with artificial tissues, functionalization, selfhealing properties, and selfbuilding robot concepts could represent the right approach for a new generation of growing robots.
Growing robots can be particularly useful in unstructured environments, especially for search and rescue applications where the tasks and constraints of the environment are not known beforehand. From a longterm perspective, it is possible to envisage a robot that could be released in its minimal configura tion and then, with materials available in the environment, could build itself: it could increase and adapt its structure to accomplish the required tasks. This would particularly benefit space and exploratory applications where growing robots would occupy less space. In addition in terms of costs, it would be a much cheaper solution to send a single unit that is able to selfbuild its structure than a set of modules.

AUTHOR CONTRiBUTiONS
ED researched the literature, discussed and wrote the paper; AS, AM, VM, and BM discussed and wrote the paper.