We use the term “problem-solving” to refer, in a broad and inclusive manner, to the kinds of things that we associate with being capable or clever. Nowadays, for many, “the ability to solve problems is not just an aspect or feature of intelligence—it is the essence of intelligence” (Hambrick et al., 2020, p.553). It is certainly not a new idea. The very birth of Artificial Intelligence owes much to it (Newell et al., 1958). As Newell and Simon later stated: “Since ability to solve problems is generally taken as a prime indicator that a system has intelligence, it is natural that much of the history of artificial intelligence is taken up with attempts to build and understand problem-solving systems” (Newell and Simon, 1976, p. 120). The task of artificial intelligence was thus to devise potential solution-generators and to design tests that could evaluate them. This assumed that the problem space was well fixed so that solutions could both be evaluated and generated. Decades later, Artificial Intelligence handbooks still devote their first central sections to problem solving (e.g., see part II of Norvig and Russell, 2021, 4th edition).

A problem is a context in which behaviors can be evaluated according to a norm of success at “solving” the problem. Problems vary from being trivial to challenging to impossible as the proportion of behaviors that are good (out of all possible behaviors) shrinks. They include “high-level” human problem-solving, such as the skills that are taught in schools or universities, as well as embodied problems such as balancing on two feet or swimming efficiently. For a system to be evaluated in terms of problem-solving, one must first have the specification of the context and of a normative evaluation so that behaviors within that context can be compared as more or less successful. Defined as such, just about anything can be seen as a problem solver. A bottle lid solves the problem of preventing spills; a car's differential solves the problem of distributing force effectively to its wheels; a computer program solves the problem of beating a human at chess.

Herein lies both the advantage and the disadvantage of placing problem solving at the center of the cognitive sciences: almost anything can be evaluated in terms of its problem solving ability. On the one hand this is a great boon. Quantifying problem solving ability is relatively straight-forward and the ability to quantify how effective a system is at solving one or more problems facilitates technoscientific progress where measurements can play an important role in defining progress. This is apparent in artificial intelligence (AI) research where benchmarks such as chess and other games (Canaan et al., 2019), hand-writing recognition (e.g., Graves and Schmidhuber, 2009), image classification, speech recognition, etc. (e.g., MLPerf benchmarks MLPerf, 2021), are used to quantify progress. Problem-solving similarly provides metrics for studying the minds in psychology and neuroscience, where problem-solving related notion of ‘tasks’ (e.g., the Simon task) are used to structure human activity and performance related metrics such as reaction speed or error rate are seen as providing key insights into how our minds operate (Simon and Wolf, 1963).

The ease of measuring problem-solving ability sometimes leads to it (problem-solving ability) becoming the explanandum—the thing we strive to understand. This is seen in research questions like: How do people recognize faces so well? How do babies come to understand the motives of other people? How do we play chess? How can we make a safe self-driving car? etc. Problem-solving also sometimes becomes the explananda—the terms in which we explain what minds are, how they work, or what they do. Evolutionary psychology (Pinker, 1997; Buss, 1998), for example, emphasizes the evolutionary advantage of problem-solving ability and in this context, explanations that are provided in terms of problem-solving ability are seen as complete, as evolution can be invoked to explain why such mechanisms exist. The evolutionary advantage of having a mind is in its contribution to problem solving and therefore, minds are best understood as problem solving machines.

However, it has been argued that: “The essence of intelligence is to act appropriately when there is no simple pre-definition of the problem or the space of states in which to search for a solution. Rational search within a problem space is not possible until the space itself has been created, and is useful only to the extent that the formal structure corresponds effectively to the situation” (Winograd and Flores, 1987, p.98). Thus, even from a problem solving perspective, intelligence is not really the capacity to solve a problem but to bring a situation into a fabricated frame where it can be treated as a problem to be solved.

Moreover, the problem with excessive focus upon problem-solving is that there are other unique and important features of minds that are worthy of study—features that may only indirectly relate to problem-solving ability or perhaps not at all. The problem, in a nutshell, is the conflation of (i) “problem-solving ability” with (ii) all of the other phenomena associated with “being a mindful body.”

The mainstream computational functionalist approach to the mind (Putnam, 1965; Fodor, 1968) doesn't really help much addressing what mindful bodies are beyond problem-solving devices. For Putnam, the very definition of the mental is always in reference to a Turing machine table that works out rational transitions (e.g., computing and storing preferences over a utility function or solving problems in problem representation space). Deviations from this rationality are treated as pathological. All human mental life is, according to Putnam, not perfectly normal, thus relatively pathological. Putnam acknowledges “our model is an overly simple and overly rationalistic one in a number of respects. However, it would be easy, in principle, although perhaps impossible in practice, to complicate our model in all these respects—to make the model dynamical, to allow for irrationalities in preference, to allow for irrationalities in the inductive logic of the machine, to allow for deviations from the rule: maximize the estimated utility. But I do not believe that any of these complications would affect the philosophical conclusions reached in this paper” (Putnam, 1965, p. 43).

Deviations from the abstract rational rule are pathological. Explanation lies on the pure domain of abstract problem solving, the deviations make it all more complicated (as if dirt in the form of a set of exceptions where to be added to the pure explanation) but change fundamentally nothing. As we are about to see, embodied approaches, and particularly enactivism, bring these “pathological” expressions to the center of the explanation (of which rational thinking is the complicated achievement) turning it into the core constitution of mind. Biologically inspired robotics has a lot to contribute in this direction.

Enactivism was born under the conviction that robotics, as a field, would require, or even force, cognitive science to move beyond the problem-solving framework:

Yet, moving away from the problem-solving paradigm has taken a long path, most of which, rather than abandoning problem-solving has deeply transformed the way we understand how nature solves those problems. In a sense, biologically inspired robotics has mostly followed the problem-solving approach and biological inspiration has focused on picking up biological mechanisms to solve problems: from the internal neuronal inspiration of artificial neural networks since its early conception (Rosenblatt, 1958) to their later development (Rumelhart et al., 1988) to the embodied strategies that either transform the problems to be solved by their “brains” or have outsourced the computational load of the problem solving to body and world (Pfeifer and Scheier, 2001). What radically distinguished biologically inspired robotics from GOFAI (Good Old Fashioned Artificial Intelligence) was a change of focus from the abstract to the concrete, from the symbolic to the sensorimotor and from the rational to the practical know-how of situated action. Despite the emphasis on self-organization, agent-environment emergence of behavioral functioning (Steels, 1990) embodiment and situated action (Maes, 1990), etc. the main goal was still to build robots capable of solving behavioral problems. After all, to put it with biologically-inspired roboticist Barbara Webb: “The sensorimotor problems faced by animals and by robots have much in common” (Webb, 1995, p. 117) and, not only can animals help us devise robots that solve problems in a biologically inspired manner, but also solving a sensorimotor problem with the robot could help us understand how the animal solves it; like “[d]etecting which ear is closer to the sound” which “is a non-trivial problem for the cricket” (Webb, 1995, p. 120).

Other trends of biological inspiration have built upon evolutionary theory itself and artificially evolved brains or brains and bodies to solve the problems (encoded as fitness function) in what is commonly known as evolutionary robotics (Cliff et al., 1993; Nolfi et al., 2016). Random variations to the parameters of robotic brain's and bodies are selected against a fitness function that operates as the benchmark of the problem to be solved. Despite the problem-solving focus that is almost inherent in artificial evolutionary optimization techniques, evolutionary robotics served to disclose a number of principles of behavioral self-organization that non-linear, fine grained agent-environment coupling display when artificial evolution can freely explore the solution space without the prejudices inherent to the human design (Harvey et al., 1997).

Some of these approaches entail radical departures from core assumptions of the computational functionalist theory of the mind: cognitive processing does not only occur in the head and the body must be integrated as a key feature of cognitive problem solving (not simply as an executioner of the solution representation worked out in the head or a sensory transmitter of the problem into it); agent-environment interactions can self-organize with little if any representations; material bodily and interactive constitution are not mere implementation details of abstract capacities but intrinsic part of the problems and solutions that cut across them.

But enactive robotics moves yet further into biological inspiration. On the one hand there are enactive approaches that have attempted to introduce more of the living body of natural intelligence into robotics by including self-organized mechanical, soft bodies (Man and Damasio, 2019) or even chemical bodies (Damiano and Stano, 2021). But also, and perhaps most relevant for enactive theory transferring to the robot what metabolism has to offer to anchor intrinsic needs or emotional feedback (Ziemke and Lowe, 2009). There is however another enactive path that brings into robotics some principles of living organization and, more specifically, the autonomy of behavior, a way of life for robots, not as somehow transferred from the biological body, but enacted at the scale of brain-body-world dynamics (Barandiaran and Moreno, 2006). Some forerunners of this inspiration are no doubt Ross Ashby on the organism-centered inspiration of adaptive controllers as machines capable to remain homeostatic in the face of perturbations (Ashby, 1952) and Grey Walter's “life imitating” robots “designed to illustrate the uncertainty, randomness, free will or independence so strikingly absent in most well-designed machines” (Grey Walter, 1950, p. 44).

More recent development of this line of inspiration on natural and biological principles for the design of robots came hand in hand with the development of a theory of autonomous behavior and agency (Smithers, 1997), organismically inspired robotics (Di Paolo, 2003), and habit-centered enactive robotics (Egbert and Barandiaran, 2014).

Varela, Thompson and Rosch opened up a new way of thinking in 1991. Their enactive approach conceives that “cognitive structures emerge from the recurrent sensorimotor patterns that enable action to be perceptually guided” (Varela et al., 1991, p.173). They later stated that: “[C]ognition is no longer seen as problem solving on the basis of representations; instead, cognition in its most encompassing sense consists in the enactment or a bringing forth of a world by a viable history of structural coupling” (Varela et al., 1991, p.205). The enactive approach thus emphasizes sensorimotor coupling and the recurrent patterns that emerge from agent-environment interactions.

This way Varela overcame the operational (en)closure of the nervous system that served as his main analogy with the organization of the living, captured (together with Humberto Maturana) within the theory of autopoiesis. Ever since, the relationship between the self-organizing nature of nervous activity and that of behavior became to some extent problematic (see Barandiaran, 2017 for a discussion).

Inspired by Maturana and Varela, Tim Smithers re-states the need for biologically inspired autonomy in robotics, in the context of the impossibility to design robot agents from a problem-solving stand point:

This analogy between metabolic autonomy and the autonomy of behavior was further explored in Di Paolo (2003). According to this view, enactivism needs not be committed to build bio-chemically living robots (provided that this is possible or even desirable) but to endow a robot with a way of life. This intuition was further explored in Barandiaran (2007, 2008).

Sensorimotor Contingency Theory can further enrich this approach. It uses regularities in the ways that motor activity affect sensory activity (sensorimotor contingencies) to explain the qualities of perceptual experience (O'Regan and Noë, 2001). Empirical research involving sensory substitution, sensory modification, psychophysics has informed the development of sensorimotor contingency theory (SMCT), which attempts to explain diverse aspects of perceptual experience, including why certain sensory modalities have a particular “feel” to them; how it is possible to make one experience one sensory modality (e.g., touch) in a way that feels more like another (e.g., sight); and the conditions in which subjects are (or are not) capable of adapting to major transformations to their sensorium. The key idea in SMCT is the role that action plays in perception: a classical enactive theme (Noë, 2004).

Using variations of the basic theme of how motor activity modifies sensory input, a set of robotic architectures where made using sensorimotor contingencies as building blocks for robotic design (Maye and Engel, 2013; Jacquey et al., 2019). But these are hardly enactive in the sense of the deep biological inspiration that the enactive approach can offer.

Perhaps the best way to explore such potential is to bring forth the concept of sensorimotor autonomy (updated and refined from a previous proposal of Mental Life and also explored in more detail on the concept of Sensorimotor Life): the capacity of an agent to sustain and regulate the structures that generate behavior. This definition echoes the metabolism-based definitions of life as far-from-thermodynamic equilibrium chemical systems capable of maintaining the network of chemical reactions that constitute it (Gánti, 1975; Maturana and Varela, 1980; Rosen, 1991; Ruiz-Mirazo et al., 2004; Luisi, 2006).

The basic constituent of sensorimotor autonomy is a sensorimotor structure (a behavioral scheme or habit) made possible by both a set embodied-neural pathways and a set of sensorimotor contingency relationships. Think of it as a coordination pattern that emerges out of environmental, sensorimotor and neural (or behavior generating) mechanisms. Now, if this structure is far-from-equilibrium or, said differently, if let alone it tends to extinguish or vanish, and if the very enactment of the sensorimotor scheme reinforces itself by repetition or by satisfying certain conditions that feed-back into its supporting structure (e.g., reinforcement of synaptic connections by Hebbian learning or reward reinforcement), then we have first sense of self-maintenance that is characteristic of habits. The more the habit is enacted the more it is strenghthened, the stronger it is the more likely it is to be enacted. That is the virtuous (or vicious!) self-sustaining nature of the habit.

We can now go back to the original, albeit obscure and self-referential, intuition of roboticist Tim Smithers and his idea of the autonomy of behavior based on the “processes that, through interaction, are continuously forming the laws of interaction that can sustain and maintain the interaction needed to form them.” These laws or norms are nothing other than the very conditions under which the habit is viable, that it can persist and sustain itself. This is a strong analogy with (metabolic) life that opens up the very possibility of having a problem of your own and having to solve it. The problem for the precarious, self-maintaining autonomous cell is persistence, avoiding decay and disintegration (see Barandiaran and Egbert, 2013 for a more detailed analysis). The same goes for the precarious, self-maintaining nature of a habit, and ultimately of autonomous sensorimotor life (see Barandiaran, 2008; Di Paolo et al., 2017). This approach also opens up a new mode of explanation that is characteristic of biological thinking and can be applied to sensorimotor dynamics: focusing on the nature and structure of constraints rather than the problems they are suppose to be adapted to solve. We return to these themes in the discussion section.

In what follows we introduce habit-based enactive robotics to illustrate and further develop the points we have briefly outlined. Inspired by SMC and enactive principles we build robots that are capable of generating spontaneously a complete ecology of habits that display structural constraints within the sensorimotor space. The results will help us discuss how enactive robotics can contribute to a new understanding of mind and cognition with a deeper biological inspiration than what problem-solving can provide.

The computational model simulates a two-wheeled robot that moves around a two-dimensional environment. The robot has two independently controlled motors; one for each of its wheels, allowing it to move forwards or backwards in a straight line or to turn in a variety of arcs, or on the spot. The robot's motors are controlled by an iterant deformable sensorimotor medium (IDSM) (Egbert and Barandiaran, 2014)—a habit-forming controller that is described in detail below.

The robot's environment is periodic in the sense that when the robot moves off one side of the environment, it appears on the opposite side. A stimulus source is also located in the environment. This source, which moves around the environment, can be thought of as simultaneously emitting a sound tone and a source of light, but in any given trial the robot is only capable of perceiving one of those sensory modalities (light or sound). We now present each element of the simulation in detail.

The robot and its environment are simulated using Euler forward integration with a time step of Δt = 0.01. Thus, a single time-unit consists of 100 iterations, and the IDSM is updated every iteration.

The stimulus source moves around the environment in a circle. Each rotation, it slows to a stop at its left-most position, before accelerating again to complete another rotation. This trajectory is specified by the following equations which describe the stimulus's position (s_x, s_y) as a function of time (t): (1)sx=12+cos(t10+sin(t10)) (2)sy=12+sin(t10+sin(t10)).

The simulated robot (Figure 1) has two independently controlled motorized wheels. It's position changes according to the following differential equations, (3)dxdt=kscos(α)(ml+mr) (4)dydt=kssin(α)(ml+mr) (5)dαdt=ks(mr-ml)2R, where x and y are the robot's position in the environment; α is its heading; k_s = 0.25 scales the speed of the motors; R = 0.05 is the robot's radius; and the variables m_l and m_r represent the velocities of the robot's left and right wheel motors.

We consider two robot sensor configurations. “Visual” robots have two directional sensors. The excitation of each, V, is the product of an attenuation factor due to distance from that sensor to the stimulus, and an attenuation factor due to misalignment between the orientation of the sensor and the relative direction of the stimulus. This second attenuation factor is calculated by taking the scalar product of a unit vector that points from the sensor to the stimulus and o→, a unit vector that specifies the direction that the sensor is facing.

Formally, (6)V=0.250.25+∥r→∥2︸distance(r→∥r→∥·o→)+︸orientation, where r→ is a vector that describes the position of the stimulus relative to the sensor; ∥r∥ is the magnitude of that vector; and the + superscript indicates that negative values within the parentheses are truncated to zero. The excitation of these sensors is thus highest when the sensor is close to the stimulus and directly facing it.

“Auditory” robots have two sensors that respond to the rate at which the sensor is approaching or moving away from the stimulus source. This is analogous to the Doppler effect whereby the perceived frequency of a sound when approaching is higher than when moving away from the listener. The excitation of an auditory sensor, A, is given by: (7)A=12+kd||r||dt where the first term can be thought of as the tones natural pitch which is offset by the relative speed of the sensor and the stimulus scaled by k=34 to keep the magnitude of the sensor (given the relative speeds of the robot and the stimulus) within a similar range of excitation values as simulated for the visual sensor.

To address the periodic boundaries of the environment, auditory sensors always use the nearest stimulus source as defined by the minimum image convention and visual sensors calculate the combined effect of 5 stimuli: one in the simulated space and four virtual copies of the sensor offset north, south, west and east of the simulated space arena by one arena width. This means that if, for example, a visual robot is close to the north boundary of the arena and facing north it can still see stimuli.

Visual sensors are offset from the direction the robot is facing by β_v = π/5 (see Figure 1). Auditory sensors are offset by β_a = π/2. The orientation of visual sensors is α ± β_v, i.e., perpendicular to the tangent of the robot's circular body at that position, facing outwards. Auditory sensors have no orientation.

The path taken by the visual robot through its environment is displayed in Figure 3A. This is a complicated and difficult to visualize trajectory, but it in fact involves several distinct repeated patterns of sensorimotor behavior (Figures 3B–H). To identify these, we first plotted the proximity of each of the IDSM's nodes to the robot's sensorimotor state over the course of the simulation (Figure 4). In this time series, it is easy to observe segments of time in which a particular set of nodes is repeatedly visited. For instance, when 750 ⪅ t ⪅ 1,200, there are a few nodes with indices close to 5,000 that are repeatedly visited (tan horizontal sequence of dots), and the same set of nodes are briefly revisited three times in the final 500 time units of the simulation. We manually identified these repeated patterns of sensorimotor activity, and assigned each pattern a label and a color (the example just provided is labeled “D” and colored tan). Times when the robots behavior is not clearly repeated were not assigned a label and are marked with a light gray color. To be clear, repetitions were identified in sensorimotor space (via the identification of repeatedly visited nodes), not in physical space. Returning to Figure 3, we can see that when the robot is performing the D pattern, it is moving around the environment, regularly turning in loops with squarish corners. Figure 5 shows how the state of the sensors, motors and the distance between the stimulus and the robot as the trial progresses. We can see that each of the colored regions tend to occupy particular regions of the sensorimotor space (the sensorimotor ‘habitat’ of the habit—see Buhrmann et al., 2013) and certain patterns of behavior involve the robot pursuing the moving light (e.g., F-behavior) while others avoid it (e.g., D-behavior).

Each of patterns B–H can be construed as a potential habit, i.e., a pattern of autonomous, self-maintaining sensorimotor activity. To justify this claim, we observe that

From these observations it follows that the nodes enable the very thing that they depend upon to persist and reflexively, the enactment of the pattern of behavior enables the very thing that it depends upon (the nodes) for its persistence. Similar to a model of biological autopoiesis (Varela et al., 1974), where inherently unstable components such as metabolism and membrane are mutually enabling, the components of these patterns of behavior in this model are in a relationship of mutual interdependence and support, and thus can be considered autonomous under some readings of the enactivist literature. The behavior itself is operationally closed entity (a “unity”), constituted by inherently unstable components (the nodes and the sensorimotor trajectory) and yet persists thanks to its own enactment or performance.

More detailed analysis of the operational closure of these systems is outside the scope of the present paper (but see Egbert, 2018 for initial analysis of the precarious autonomy of a simple IDSM-based habit). We can see the basic idea however, when we consider how the weight of the nodes changes as time passes in the simulation (see Figure 6). Recalling that when the weight of a node reaches zero, that node ceases to exist, it is clear that only nodes that are regularly visited can persist in the long term. For a collection of nodes to be visited, the sensorimotor trajectory must move in a particular trajectory and the sensorimotor trajectory is largely determined by the activity of nodes. It follows from this that the only way that a pattern of behavior can persist is if it is one that causes the repeated revisitation of its constituent nodes. We can also note that this self-reinforcement of a habit need not be constant or contiguous. For example, the B-habit (red) is established early in the simulation [t ≈ 250)] and then is not visited again until t ≈ 1,250, where it is enacted a few times and the nodes are reinforced such that they survive until close to the end of the simulation. The parameters that prescribe the rates of node weight reinforcement (k_↑) and degradation (k_↓) determine how regularly patterns of behavior must be enacted if they are to persist.

We performed a similar analysis on the auditory robot. Figure 7A shows the complete path taken by the auditory robot in its environment, broken down by habit (Figures 7B–H). Figure 8 shows the weight of each node which also provides an indication for how close each node is to the current sensorimotor state of the robot and the timeseries plots for the auditory robot are shown in Figure 9. It is worth noting that though these patterns of behavior may seem random, they all (both visual and auditory) relate to the moving stimulus source in regular, non-random ways. The regularities in the interaction with the stimulus are much more easily seen in the animations linked to in the captions of Figures 3, 7).

Even when two behavior's functions are the same (e.g., they both accomplish taxis) the forms of the underlying sensorimotor habit can be different. As a case in point, we can compare auditory habit E with visual habit F. In both cases, the robot circles around the stimulus, maintaining a similar approximate distance from the stimulus as the stimulus moves in spurts around the environment. When we look at the two patterns of sensorimotor activity (Figure 10), we see two different pictures. Each row of this visualization shows the spatial trajectory of the robot (left) followed by four projections of the 4D sensorimotor trajectory of the robot and it is obvious that the two sensorimotor trajectories are qualitatively different. The auditory pattern involves less diversity in the range of states visited; the mean and other statistical properties of the sensorimotor trajectories are also clearly different, etc. But every habit in this model is unique—how much of the difference between these habits is simply due to the fact that they are different habits, and how much is due to the difference in the robot's sensorimotor modalities?

To address this question, we can look at projections of the sensorimotor trajectories of all of the habits of the visual robot (Figure 13) and observe that there are regularities in the forms of the habits of this robot. This is most readily seen in the left-sensor vs. right-sensor projection (upper right) which reveals elements of this robot's sensorimotor contingencies, including aspects of both its sensorimotor habitat and environment (Buhrmann et al., 2013) of the robot. Specifically, the bilateral symmetry of the robot's embodiment result in a (statistical) mirror symmetry across the left sensor = right sensor diagonal. The directionality of the visual sensors and their different orientations mean that it is impossible (in the environment with a single stimulus) to maximally stimulate both left and right sensors concurrently, leading to the L-shape apparent in upper-right plot of Figure 13, with arcs that pass between the stems of the L as the robot turns such that the stimulus moves from being in front of one sensor to in front of the other. These features are also found in the other nine visual robot trials.

Similar plots for the auditory robots (Figure 14) also reveal regularities. Like the visual robot, the auditory robot's embodiment is bilaterally symmetric and so it has the same symmetry across the diagonal in left-sensor/right-sensor plot, but the non-directionality of the auditory sensors and the fact that their excitation does not fall off with distance make other aspects of the sensorimotor activity different. The auditory trajectories tend to repeat patterns of diverging from a center point in short arcs in a pattern of returning and diverging again, in a different direction. These dynamics can be seen in greater detail in Appendix A, which shows the projection in sensorimotor space of each habit individually.

Sensors and motors are involved in a recursive relationship of influence, where sensors influence motors (via the controller) and motors influence sensors (via the effect of actions upon the environment and the agent's relationship to the environment). It therefore would make sense that the constraints imposed upon sensory dynamics by different sensory modalities would constrain the motor dynamics. In other words, under one sensory modality, certain motor trajectories will be more readily repeated (and thus stabilized) than others, and which motor trajectories are more readily repeated would depend upon the sensory modality.

We do, in fact, see differences in the distributions of motor states between the two sensory conditions (Figure 11). Visually, we can see that the distributions of motor (and abstract motor) states in the visual robots seem different than that of the auditory robots (Figure 11); and the distributions of motor states within these groups in independent runs of the visual trial seem more similar to each other (Figure 12) than they are to the other group. We confirmed the statistical significance of these differences by using Kullback–Leibler divergence to assess the distributional conformity of 50 visual trials and 50 auditory trials. The within-group (intra-modal) distributions were significantly more similar than the between-group (inter-modal) distributions for m (t = –4.67, p = 0.0114) and for μ (t = –8.507, p = 0.0215).

In the absence of the self-reinforcing dynamic of the IDSM, habits, motor activity would be random, in which case we would expect the same distribution of motor states in both sensory modalities. The histograms in (new) (Figures 11, 12) show that this is not the case, and thus provide support for the claim that the sensorimotor contingencies implicit in the different sensory modalities constrain not only the sensory dynamics, but also motor dynamics—and in doing so, they play an important role of constraining the form of emergent autonomous sensorimotor dynamics.

The primary point that we wish to communicate using this model is that sensorimotor contingencies influence which patterns of autonomous sensorimotor behavior can emerge and persist. It is worth emphasizing that the plots in Figures 10, 13, 14 are not plots of the sensorimotor contingencies of the robot—they are plots of the robots' self-maintaining habits. Nevertheless, we see in these plots the influence of sensorimotor contingencies—the way that the contingencies have constrained the set of habits that can form and self-stabilize, producing the regular patterns described in the paragraphs above.

The two different sensory modalities entail two different sets of sensorimotor contingencies. The different sensorimotor contingencies each imply a different set of possible self-stabilizing habits, and so the habits that form with one sensory modality are more similar to each other than they are to the habits that form with the other sensory modality. Two visual habits will be different from each other, but they will generally be more similar to each other than auditory habits. This is what we mean by the claim that the form of sensorimotor habits are constrained by their sensory modality and it is apparent in the regularities described above and seen in Figures 13, 14. In this account, sensorimotor contingencies play an important role. They influence the form of the emergent autonomous sensorimotor strutures, but the rules themselves are not internalized or learned in any other way.

The robots we have just presented display a number of interesting and challenging features for traditional problem-solving centered artificial intelligence (biologically inspired or otherwise). We see in this model how sensorimotor dynamics can have problems of their own. As we have seen AI and robots more specifically are generally built to solve problems. But whose problems? Perhaps the engineers' problems or those of their clients, perhaps animal problems that require robotic models to be better understood, maybe societies problems but certainly not those of the robot itself. It is thus important to distinguish between having one's problem vs. being able to solve a problem. A bacterium that swims up a gradient toward the resources it needs to survive is doing both. It has a problem (it needs food to persist) and it solves that problem by navigating through its environment. AI, on the other hand, is capable only of the former. A “self-driving” car that takes me to a restaurant dinner is solving a problem (my need for food), but that problem is mine, and not the car's.

The enactive robots presented in this paper where not designed to do anything, to solve any problem. None of our analysis was described in terms of success or failure at the performance of a task. The architecture makes habits emerge and with them a basic sense of having a problem: the habit needs to enact itself in order to persist. The norms of behavior here are thus only in terms of the persistence of the behavior itself (and concurrently the mechanism that produces it). Surely, life is full of problems, we certainly have enough with our own and don't need to be concerned with creating new ones …unless we want to understand what it is to have a problem and how life is, itself, a source of problems. Enactivism embraces a deep conception of life as the self maintenance of a network of precarious processes. In this sense life is inherently a source of problems. By making it possible for habits to emerge and sustain their own existence through recurrent sensorimotor interaction we have shown how this conception of life can be practically transferred to a “way of life” in sensorimotor robotics.

A research program on enactive robotics can and must address how sensorimotor life is a source of problems. We can move beyond the single habit and envision how the network of habits that constitutes the agents can display emergent problems of its own: that of keeping the whole network alive, coping with variations on the way the environment affords or precludes to enact them. This line of enquiry requires further work but the present model requires little development to start addressing it.

Biological inspiration in robotics can also move beyond the analogy between metabolic or physiological life and its sensorimotor counterpart. In particular the models presented and analyzed in this paper show how sensorimotor life and the problems that emerge within are constrained by sensory embodiment and how such constraints constitute an important part of what it means to explain life.

Whether it is understood as more fundamental-than, complementary-to or directly at odds-with (Newman, 2018) natural selection, it is the repeated, regular, robust and often phylogenetically independent appearance of forms or structures in biology what becomes in itself an object of study that cannot be reduced-to or deduced-from life as a response to environmental problems. In this sense biologically inspired enactive robotics can learn from biological development and its organization and import into psychology some explanatory strategies followed by biologists. Darcy Thompson, René Thom, Pere Albert, etc. all conceived biological forms to be both explananda and explanans on themselves. Homologies, analogies, inherency, are all concepts directed to capture the re-emergence of certain forms (and functions!) in biological organisms. The notion of constraint plays here a fundamental role. The way in which different layers of materiality and self-organization limit and channel the emergence of viable forms is essential to biological explanation.

If biological inspiration within the problem-solving paradigm concluded that the material embodiment of robots permits to offload and transform the problems to be solved, the biological inspiration of the enactive paradigm can conclude that the embodiment loads and informs the problems that constitute the sensorimotor agent. As we have seen, sensory modalities constraint motor trajectories, which in turn, shape how habits get stabilized. Thus, in a way that parallels the explanatory role of constrains in evolution, we can hypothesize that behavioral variability is not free (to explore potential solutions to cognitive problems) but is constrained and channeled by the embodiment of the agent.

Now, before the advent and widespread influence of evolutionary theory in biology, the latter synthesis with molecular biology and subsequent expansion to psychology, sociology and even epistemology, or philosophy more generally, the concept that was key for the continuity between life and mind was that of habit.

There are different takes on how enactivism translates into robotics. As mentioned in the introduction it is possible to simply reject representational functionalism, or to claim for the importance of embodiment, or to demand that robot be endowed with living bodies, or to introduce some feedback mechanisms that parallel those provided by emotion bearing bodies. We have taken a different approach—that envisioned by Smithers (1997) and latter developed by Di Paolo (2003), Barandiaran and Moreno (2006), and Barandiaran (2008). What a research programme in enactive robotics entails is the study of the organization of sensorimotor life: the form of viability of different habits, the topology of the network of habits that unfolds over development, the shape of the habitat that is thus constructed, the structure of the world that is experienced. Habit-based robots as designed here are capable of individuating habits and of creating an ecology of habits that can easily be understood as a form of sensorimotor self.

Not only has enactivism informed robotics (Ziemke and Lowe, 2009; Vernon, 2010) but robotics has often served enactivism (Beer, 1995; Di Paolo, 2003; Aguilera et al., 2016) by clarifying its claims, pushing theoretical development, operationalizing its concepts or penetrating diverse problems. The model we developed here can be aligned with the latter. It can be used to clarify and make explicit the often obscure original formulation of enactivism that “cognition in its most encompassing sense consists in the enactment or bringing forth of a world by a viable history of structural coupling” (Varela et al., 1991, p. 205). We have shown how a robot endowed with an IDSM can bring into being a number of habits, the world that it brings forth is the habitat, more specifically, the structured set of sensorimotor contingencies that the agent inhabits or enacts. This habitat must be viable in the sense that habits must be sustainable and results from a (developmental) history of sensorimotor (structural) coupling. And in so doing, we have demonstrated how sensorimotor contingencies can directly constrain or “sculpt” the form taken by sensorimotor habits without requiring any virtualization, i.e., without any internal model or representation of the sensorimotor contingency.

Perhaps it is worth clarifying why the nodes do not constitute an internal model or representation. The nodes do not stand-for something else other than the sensorimotor dynamics they partake in. It is not possible to operate upon the nodes in a decoupled offline mode that is not itself the enactment of a behavior, and there is not additional module or subsystem “consuming” such nodes to carry out any further operation. In these ways, the nodes do not represent a behavior or habit, but they embody it—they constitute the habit together with elements of the robot's bodily and environmental dynamics.