Building upon the work done in the previous section, we claim that the challenge for a radical embodied theory of skilled motor behavior is to explain intelligent, cognitive control over action without assuming the existence of top-down representational processes. In this section, we argue that ecological psychology (Gibson, 1979; Chemero, 2009; Turvey, 2018; Segundo-Ortin et al., 2019) offers valuable conceptual, empirical, and methodological resources to start facing this challenge.
An ecological approach to skilled action
A foundational idea of ecological psychology is that perception is direct. To say that perception is direct is to say that perception is not mediated by internal inferences or computations. Rather, perception consists of the detection (or “picking up”) of perceptual information. To better understand this idea, we shall elaborate upon the notion of “perceptual information.”
For ecological psychology, perceptual information is contingent on the existence of spatial-temporally extended and structured patterns in the topology of the ambient energy array (Lobo et al., 2018). Imagine a room illuminated by a ceiling light. As the light interacts with the surface of the objects that furnish the room, it generates a specific pattern—an ambient optic array. This pattern is not random or stochastic. Rather, it is lawfully generated given the conditions of the light and the objects—their situation, orientation towards the source of light, but also the materials they are made of. Ecological psychologists refer to the lawful relation that exists between the environmental properties and the ambient energy distribution as “specificity.” To say that the ambient optic array is specific to the surfaces of the room is to say that there is a one-to-one covariation between the former and the latter.Footnote 8 Because a particular surface α in the room lawfully generates a specific pattern β in the ambient optic array, the occurrence of β guarantees the presence of α.
To fully appreciate the importance of specificity, it is worth remembering that the starting point of traditional (representational) theories of perception is the inability of light (or sound, or chemicals in the air, etc.) to specify their environmental causes. Because there is a many-to-one mapping of ambient patterns to environmental causes, perception necessitates “an almost unimaginably difficult causal inference problem: [brains] must infer the hidden state of the constantly changing environment from its profoundly non-linear and ambiguous effects on the organism’s numerous sensory transducers” (Williams, 2018, p. 150). The necessity to explain this causal inference calls for a representational framework.
By contrast, the hypothesis of specificity denies that perceptual information is necessarily ambiguous and impoverished, thus parting with the necessity of a representational explanation for perception: If β specifies α, then detecting (paying attention to) β suffices to perceive α:
Direct perception can be possible if properties of the world are specified in patterns of stimulus energy. If patterns of the world are unambiguously specified, perception does not have to involve processes of interpreting ambiguous cues. (Fajen et al., 2009, p. 81)
Thus, agents can be aware of the different properties of the performing situation by detecting information in the ambient array, but this awareness is not the product of mediating inferences and mental representations.Footnote 9
Another foundational idea of the ecological approach is that perception is primarily for action. According to J. J. Gibson (1979), because the properties of the environmental setting are specified in the structure of ambient energy array, the agent can perceive what this setting affords to her by detecting this information. This implies that behavioral control is possible on the basis of direct perception, and that complex behavioral solutions that seem to require motor representations can be simplified by capitalizing on affordance-specific perceptual information:
[R]esearch shows that there is an alternative to appealing to such computational-representational structures and processes. With the right kind of information, an individual can be coupled to her task environment in a way that supports behaviour about forthcoming events without explicit predictions. (Silva et al., 2019, pp. 55-56)
For illustration, consider the perceptual-motor problem of breaking the car before crashing an approaching obstacle. Knowing the time remaining until physical contact with the obstacle is crucial for doing this. The question is how we can know this time. One possible solution is to assume that we perceive several properties of the environment—say, the current distance to the obstacle, the speed to which we approach it, etc.—and somehow combine them to infer this time. The ecological answer is radically different. According to General Tau Theory (see Lee, 2009), as the distance between the object and the visual system reduces, the object “expands” on the visual field, and the rate to which the object expands—a variable referred to as “tau” (τ)—specifies the time remaining until contact.Footnote 10 Because the object’s rate of expansion specifies time-to-contact, it provides the information needed to control action.
The empirical evidence speaks in favor of the ecological hypothesis. For instance, Fajen and Devaney (2006) successfully predicted that drivers stopping at designated points continuously adjust their breaking to keep τx(t)—this is, the rate of change of τ over time—below 0.5, the critical point at which avoiding collision is no longer afforded. After having compared different hypotheses, Fajen and Devaney concluded that drivers adjust braking on a moment-by-moment basis by detecting τx(t). Braking is thus controlled by paying attention to τx(t) without the necessity of forming and manipulating a mental model of the situation.
Strikingly, these studies also show that whereas perceivers are bad at estimating the distance to the object and the speed at which it approaches, they can successfully tell the moment at which avoiding collision is no longer possible. This supports the hypothesis that the primary objects of perception are the affordances and shows how perception can guide action without mediating inferences. If all the information needed to control action—that is, the information needed to know whether avoiding collision is possible at this rate of deceleration—is already present in the ambient optic array, then calling upon representations to explain action control is unnecessary.Footnote 11 Cognitive control over action can be achieved on the basis of direct perception.
Perceptual learning for skilled action
If successful perception–action coupling depends on detecting perceptual information, then it makes sense to conclude that “the strength of the coupling depends on the usefulness of the variable detected” (Withagen & Chemero, 2012, p. 532). This claim invites the view that perception must be trained and learned.
Along with the study of perceptual processes, another branch of ecological psychology, mostly led by Eleanor Gibson (Gibson, 1969; Gibson & Pick, 2000), has focused on studying how perception–action is trained—i.e., how perceivers get to detect new and better informational variables, thus improving overall performance. Ecological psychologists challenge the assumption that perceptual, or rather perceptual-motor, learning requires increasingly refined representations, and propose, instead, that it consists of attuning to more specific informational patterns. Therefore, according to ecological psychologists, the main factors that explain the difference between novices’ and experts’ performances in a particular task are the informational variables they use for guiding their action.
Perceptual learning is primarily depicted as involving the “education of attention.” Attention is said to be optimally educated when the perceiver comes to detect the most useful (the most specific) informational variable for the task being faced. Once the right information is detected, the perceiver must calibrate her behavior to achieve the most efficient perception–action coupling—for example, by modifying the force applied to the braking pedal in response to τx(t).Footnote 12
Jacobs and Michaels (2007; see also Higueras-Herbada et al., 2019) have depicted perceptual learning as a direct process akin to perception. The theory of direct learning rests on the hypothesis that there is information in ambient energy arrays that is specific to the possibility of reducing non-optimalities in perception–action cycles. Jacobs and Michaels refer to this type of information as “information for learning” and claim that this information is present in the observable changes of the ambient array the learner produces when she acts. According to this theory, perceptual feedback following action provides information for learning, and detecting this information is tantamount to perceive how to improve the performance in successive trials.Footnote 13
The theory of direct learning has been successfully applied to investigate the learning of different skills, from simple ones such as keeping the balance of an inverted pendulum attached to a moving cart (Jacobs et al., 2012) or the identification of different object properties by dynamic touch while blindfolded (Jacobs et al., 2009), to more difficult ones, including practicing the final approach phase in landing via a flight simulator (Huet et al., 2011). All these studies show that, after practicing with sensory feedback, novices and experts tend to converge towards the same informational variables and use these variables to improve their performance. This suggests that “the assembly of functional actions in skilled performance is a dynamical process, dependent on relevant sources of perceptual information” (Davids et al. 2013, p. 24).
Hence, whereas traditional ecological studies have focused on formulating laws that connect perception (information for affordances) with action in specific tasks, detailing “which features of the action environment the [expert] agent is attending and why” (Fridland, 2014, p. 2736), studies on perceptual learning explain how agents can improve their perception–action in these tasks without assuming they must build increasingly sophisticated internal models of the environment. The experimental evidence gathered by ecological psychologists supports the idea that novice participants become experts as they educate their attention towards more specific information (see Higueras-Herbada et al., 2019 and Lobo et al., 2019 for a review). The learning curve predicted in task-specific direct learning studies is critical to explain how perceptual learning occurs in a non-representational way.
Directly connected to the issue of perceptual learning is decision-making. Decision-making is a clear example of agentive control, and, as such, it is commonly assumed to depend on mental inferences and represented action plans (see Fridland, 2017a, 2017b). The Gibsonian literature suggests a different picture. For instance, Araujo and colleagues (Araújo et al., 2006, 2019) have applied a broadly Gibsonian framework to study decision-making in sport performance. Their approach is based on previous work by Warren (2006). Warren developed a general explanatory framework for behavioral dynamics, articulated in terms of the emergence of behavior trajectories from informational, bio-mechanical, and task constraints. In this model, agent and environment are conceived of as a pair of dynamical systems coupled by perceptual information. According to it, the confluence of the biomechanics of the (neural and non-neural) body, the structure and physics of the environment, and the available perceptual information, all of them intertwined under the boundary conditions of a particular task or goal, give rise to adaptive, intelligent behavior. Decision-making—this is, changing from one type of action to another—is thus based on the continuous exploration and detection of competing variables in the information space of the task.
As Raab and Araujo (2019) comment, when Warren’s behavioral dynamics is applied to studying sports, we find out that “decision-making emerges as athletes search in a field of affordances to arrive at a stable, functional solution” (p. 6). This idea is exemplified by Correia et al. (2012). The authors used a VR environment to simulate a 3 vs. 3 rugby task where participants had to decide between running ahead with the ball, making a short pass to a teammate, or making a long pass. Participants were exposed to different scenarios, with the most important variable being the initial distance between the opponents, themselves, and the teammates. Experimenters found that participants used information regarding the opening paths to decide how to act, and that the emerging gaps relative to the defenders, and between the defenders and the teammates, were the best predictor of the carrier’s behavior. In conditions where no gap was available, the participants most common solution was to keep the ball and wait until a sufficiently large gap emerged—that is, they waited until a better affordance appeared. For the authors, this suggests that “the action most often selected for each gap location was the affordance that was best aligned with the task goals” (p. 317). Besides, the researchers found out that “professional players were better able to distinguish the information specifying the affordance in each of the varying gap conditions” (p. 318).Footnote 14
What these empirical findings suggest is that decision-making is a dynamical process influenced by the availability and detection of affordances. Decisions occur at so-called “bifurcation points” when new information appears, and ultimately depend on the capacity of the athlete to detect this information. Decisions are made by means of perceiving better affordances. Moreover, because experts can detect more specific informational variables than the novices, then they have a higher capacity to appropriately modify their actions according to the circumstances. This explains why behavioral flexibility increases with perceptual skill, and shows that perceptual learning is crucial for improving decision-making in skilled performance.
Considering different objections
Before concluding this section, there are two potential objections that deserve consideration. To begin with, it can be argued that some claims by Gibsonian psychologists invite the reading that skilled is passive. For instance, Correia et al. (2012) assert that the information available at each stage of the task “is shaping the emerging actions” (p. 317). Likewise, Withagen et al. (2012) have argued that agents can be attracted or repelled by the affordances of the environment to act in specific ways. At first glance, this can possess a problem for us, for it invites the reading that action are drawn by the affordances and so that it is not under the agent’s control.
We argue that this is a too simplistic reading of the situation (see Segundo-Ortin, 2020). In particular, such a reading overlooks the fact that for affordances to attract or repel action they have to be perceived, and perception is in itself an active process according to ecological psychology. In fact, in The ecological approach to visual perception, J. J. Gibson (1979) urged for a redefinition of perception as “an achievement of the individual” (p. 228) and often spoke about the “act of perceiving” (p. 130).Footnote 15 For one thing, he depicted perception as an act information pickup, not a response to a stimulus. This implies that to perceive an affordance, the observer must modulate her attention, selecting (“picking up”) the informational variables that are relevant for the goal she aims to achieve. Moreover, perceptual acts often involve more than shifting our attention focus; they require skillfully exploring and manipulating the environment in order to generate the appropriate information.Footnote 16 Both kinds of exploratory activity are under the control of the agent and are necessary in order to perceive affordance. Therefore, even if affordances were shown to invite, solicit, or attract action in any way (but see Heras-Escribano, 2019a, 2019b, p. 111 for a critique of this hypothesis), this would not imply that the whole perception–action process is passive.
The idea that perception–action is active and goal-oriented invites the second worry. For representationalists, it is not clear how we can explain the goal-directed character of perception–action without assuming that goals are represented in the agent’s mind and influence, in a top-down fashion, perception and action. Before we deal with this objection, it is noteworthy that neither Fridland (2017a, 2017b) nor Bermudez (2017) offer reasons as to why goal-directedness requires discrete goal-representations. Instead, they simply assume this to be the case and focus their analysis on determining the format these representations must take. Moreover, they do not even explain how such goal-representations intervene in controlling attention and motor processes.
Fortunately, this is not the only option available. Defenders of radical embodied theories of cognition have offered different strategies to account for goal-directedness without assuming that there must be a part of the system—the goal-representation—that controls and shapes the system’s overall behavior (see Brancazio & Segundo-Ortin, 2020; Di Paolo et al., 2017; Reed, 1996). We will focus here on a proposal first articulated by Juarrero (1999) and later explored by Van Orden and colleagues (Van Orden et al., 2003) in the context of ecological psychology. The core idea is that intentional actions are not the products of discrete, causally efficacious mental states. Instead, they emanate from the self-organizing dynamics of open dynamical systems that span the brain, body, and environment. Open systems exchange matter and energy with their environment through a process of feedback loop. Thanks to this positive feedback loop, the agent affects the environment and is affected by it, shaping a common dynamical trajectory. This feedback loop gives rise to the sort of self-control that, according to Juarrero, is characteristic of purposive, goal-oriented behavior.
According to Juarrero’s view, committing to one goal does not involve forming a discrete mental representation. Instead, it implies entering a specific self-organized neural and non-neural configuration that, in turn, sets the whole system, at different physical and temporal scales, near a critical state. Critical states are those in which the system is acutely sensitive to the contextual factors that are directly relevant to the consecution of the goal, making some motor behaviors more likely to occur than others. On this view, goal-directed actions are to be understood as “dynamical trajectories” (Juarrero, 1999, p. 150) instead of as the product of discrete mental representations.
Ecological psychologists understand the previously mentioned feedback loop in terms of perception and action, and take the information about affordances as the main currency in this exchange:
Action changes the circumstances of the mind and body, which change the opportunities for perception […]. Changing propensities for action introduce new opportunities for perception. New opportunities for perception entail new propensities for action and reconfigure intentional contents. This interplay among self-organizing intentional contents and perceptually changing circumstances uniquely situates ordinary purposive behavior. (Van Orden et al., 2003, p. 333)
This idea fits well with what the ecological literature about decision-making in sport shows (see, e.g., Correia et al., 2012). Being tasked with not losing the ball to the opponent makes participants more acutely sensitive to the affordances relevant to this task. However, as the situation changes, new affordances emerge, biasing the behavior of participants toward other, more specific goals—namely, running forward instead of passing the ball to a teammate. Importantly, the intention to run forward was not there before the situation changed, but emerged once the agent perceived that passing the ball to a teammate was no longer possible. Therefore, acting intentionally does not require keeping a goal in mind, but exploiting the goal-specific information present in the ambient array.
Moreover, we believe that adopting this dynamical approach helps us make sense of recent empirical discoveries concerning the neural underpinnings of action control. Uithol et al. (2014) conducted a review of different studies concerning the activity of the lateral prefrontal cortex (lPFC)—the area of the brain that is generally recognized to be responsible for engendering and controlling action—while participants were engaged in different motor control tasks. According to the authors, all the evidence shows that the activity in the lPFC is highly-context sensitive, making it impossible to track one specific state or firing pattern that remains stable, guiding action from beginning to end:
[I]n contradiction to the condition of positing discrete representations, the activation of the neurons [within the lPFC] do in fact vary continuously during the period of action control, and moreover in a way that is sensitive to the fine-grained temporal progression of the action sequence. (Uithol et al., 2014, p. 133)
For the authors, the fact that the patterns of activation in the lPFC are continuously modified through action execution “speaks against the idea of a discrete state that begins the action coordination process, and retain its functional identity and content through an action episode” (p. 133; see also Kalis, 2019). This, the authors conclude, undermines the idea that goal-oriented behavior involves the top-down influence of goal-representations over motor control processes. For them, “the idea of a discrete intention causing and controlling actions from the top of a representational hierarchy is a mischaracterization of the complex and dynamic nature of action control” (p. 136).
Advancing a thorough account of purposive behavior in non-representational terms is beyond the scope of this paper. Instead, we want to argue that the fact that there are alternatives to representational accounts of intentional or goal-directed behavior, united to the traditional problems associated with the idea that representational contents can cause motor behavior, and the previously mentioned difficulties in finding neural states that can be identified with goal-representations, show that the view that goal-representations are required for goal-directedness is far from evident. Therefore, an argument must be forthcoming on the representationalists’ side before we can accept their claim as a matter of fact.