There is increasing consensus in cognitive science and neuroscience that we need a novel view of the brain that is more action-oriented; in this perspective, it has been argued that we might be facing a “pragmatic turn” in cognitive science (Engel et al. Reference Engel, Maye, Kurthen and König2013). One central proposal of action-oriented views is that the brain evolved to be a control system for the interaction with the external environment. We should look at cognitive (and brain) functions, including the most advanced (or “higher-cognitive”) functions, within an interacted and control-theoretic framework – as activities that an organism performs in interaction with its environment, rather than in terms of modular computational operations over discrete symbols independent from perception and action systems. This idea can be traced back to early theories in cybernetics, pragmatism, and ecological psychology (Ashby Reference Ashby1952; Craik Reference Craik1943; Gibson Reference Gibson, Shaw and Bransford1977; Wiener Reference Wiener1948) and has been often reproposed, in slightly different forms, in disciplines such as cognitive science, neuroscience, robotics, and philosophy (Cisek Reference Cisek1999; Cisek & Kalaska Reference Cisek and Kalaska2010; Clark Reference Clark1998; Engel et al. Reference Engel, Maye, Kurthen and König2013; Pezzulo Reference Pezzulo2011; Pezzulo & Castelfranchi Reference Pezzulo and Castelfranchi2009; Pezzulo et al. Reference Pezzulo, Rigoli and Friston2015; Pfeifer & Scheier Reference Pfeifer and Scheier1999; Scott Reference Scott2012; Stoianov et al. Reference Stoianov, Genovesio and Pezzulo2016; Varela et al. Reference Varela, Thompson and Rosch1992). Anderson (Reference Anderson2014) contributes to this debate both theoretically and empirically. He proposes a view of brain organization in which regions form functional coalitions dynamically and implement control loops for agent–environment interactions. He also discusses how neuronal coalitions could be reused to implement higher cognitive skills, forming possible control loops to support mathematical and language processing.

Despite its appeal, the impact of the action-oriented view is mitigated by various factors. One reason is the level of ambition of the novel proposal, which requires a unitary perspective on brain function and behavior. It is easier to study the brain if one assumes that each area or network is the locus of a single (or a small set of) cognitive operation(s) such as perception, memory, language, and so forth, and that each of these functions can be largely studied in isolation. But this perspective is not fully compatible with an action-oriented view, which assumes that the function of the brain as a whole is engaging in interactive control loops. Because these loops extend beyond the brain, we should also consider the contribution of the body and the external environment, including other agents, to cognition.

Given this level of ambition, advocating a generic “control metaphor” for cognition that replaces the older metaphors of the brain – the brain as a computer or as a serial transducer from stimuli to behavior – might be insufficient. What might be more effective is the realization of a coherent set of computationally specified process models of action-oriented cognitive processing – from sensorimotor transformations and situated choice to higher cognitive skills such as mathematics, language processing, and problem solving – that can guide empirical research and provide a more compelling and unifying explanation of empirical findings.

Indeed, despite noteworthy empirical demonstrations of the relevance of action-based concepts to understand cognition, including higher cognition (Barsalou Reference Barsalou2008; Glenberg Reference Glenberg1997; Jeannerod Reference Jeannerod2006), we still largely lack detailed process models. For example, several studies illustrate the importance of bodily processes and visuomotor strategies for decision making or problem solving (Spivey Reference Spivey2007), but the (relative) lack of mechanistic models hinders a complete understanding of these phenomena. Progress in the field will benefit from a “new alliance” between proponents of embodied and action-oriented views cognition and computational modelers including roboticists (Pezzulo et al. Reference Pezzulo, Barsalou, Cangelosi, Fischer, McRae and Spivey2011; Reference Pezzulo, Barsalou, Cangelosi, Fischer, McRae and Spivey2013). The idea that cognitive robotics can shed light on brain and cognition is relatively new, but it should not be too surprising if one considers that robots are well suited to implement the kind of brain-body-environment (and social) interactions that are considered essential for action-based cognition (Verschure et al. Reference Verschure, Pennartz and Pezzulo2014). Furthermore, developing complete models (robotic or nonrobotic) of cognitive operations can counteract an excess of specialization in the field – say, the focus on the functioning of just one brain area.

If we consider the theoretical debate within action-based theories, it often revolves around (the need for) internal representation. Various competing proposals include: abandoning the notion of internal representation to see cognition as online control and interaction with the environment (Chemero Reference Chemero2009; Gallagher Reference Gallagher2005); externalizing representation (Clark & Chalmers Reference Clark and Chalmers1998; Kirsh Reference Kirsh2010); replacing it with other constructs such as sensorimotor contingencies (O'Regan & Noe Reference O'Regan and Noe2001) or dispositions (Engel et al. Reference Engel, Maye, Kurthen and König2013); amending it in a more embodied and action-oriented view, in terms of, for example, perceptual symbols (Barsalou Reference Barsalou1999), action-oriented representations, or emulators that support control structures (Clark & Grush Reference Clark and Grush1999; Grush Reference Grush2004), to mention just a few.

Although important, this debate might be limited if not accompanied by the realization of mechanistic models, which answer specific questions such as: What kind of control system is the brain? What control loops are instantiated between the brain, the body, the environment, and with other agents, during the implementation of tasks of various complexity, such as walking, solving a puzzle, or planning a trip together? How can these and other tasks be mechanistically described in terms of agent–environment interactions, and which variables, if any, are technically “controlled” in these tasks? What is the contribution of different brain areas to each of the control loops required by different tasks, how are their contributions dynamically assembled, and which are their neuronal signatures? Which loops require a continuous engagement with external variables and which ones their internal, endogenous generation? Which aspects of the structure of the agent–environment interactions can be exploited online, which can be offloaded externally, and which, if any, need to be internalized? How can we better simulate, measure, and empirically study these interactive loops? And so on.

Several researchers are already addressing these and other relevant questions. However, the space of possibilities has not been systematically explored, with many interesting constructs from control theory, cybernetics, and related fields that remain untested. These include early proposals such as hierarchical perceptual control (Powers Reference Powers1973), the test-operate-test-exit architecture (Rosenblueth et al. Reference Rosenblueth, Wiener and Bigelow1943), and several other constructs from early cybernetics (Ashby Reference Ashby1952; Wiener Reference Wiener1948), as well as more recent developments such as, for example, optimal feedback control, model-based control, and risk-sensitive control, which are often mentioned in theoretical debates but not systematically tested (Shadmehr & Mussa-Ivaldi Reference Shadmehr and Mussa-Ivaldi2012).

Incorporating these ideas into specific process models could guide progress in psychology and neuroscience and illustrate where exactly an action-based theory differs from traditional ones. One interesting success case is the affordance competition hypothesis (Cisek Reference Cisek2006), in which action specification and selection proceed in parallel and compete – on the basis of the various sensory, contextual, affective, and motivational biases they continuously receive – until one action pattern is selected, hence dispensing from a central decision module that processes abstract values prior to action. Another recent example is the active inference framework (Friston et al. Reference Friston, Daunizeau, Kilner and Kiebel2010; Reference Friston, Rigoli, Ognibene, Mathys, FitzGerald and Pezzulo2015). It can be considered a modern development of cybernetic ideas (Seth 2014), although it casts control as a (probabilistic) inference problem (see also Attias Reference Attias, Bishop and Frey2003; Toussaint Reference Toussaint2009). It challenges traditional models of perception and action control in several ways – for example, by natively incorporating top-down processes and by highlighting the importance of predictive processing and error correction over and above stimulus-response associations – and it paves the way to the realization of models of more elaborated cognitive operations, suggesting that they might be based on the same principles (Clark Reference Clark2013b; Pezzulo Reference Pezzulo2012).

Developing mechanistic models can help advance our theoretical understanding of action-based cognition, too, because some of the aforementioned models cut across dichotomies (e.g., interactivity vs. internal modeling) that are often debated, including in Anderson (Reference Anderson2014). For example, model-based control and active inference offer a dynamical view of online action control and interaction but include internal generative and predictive modeling loops to make those interactions more effective. These systems learn the statistical structure of the environment and of agent–environment interactions, and they perform state estimation. However, structure learning and estimation are functional to effective control rather than having as a final goal a veridical representation of the external environment. These models remain to be fully tested. However, these examples suggest that the development and empirical testing of specific process models might contribute to the theoretical debate on action-based cognition – for example, by clarifying the possible contribution of predictive processing and structure learning to control and cognition.

These arguments are also important to meet another challenge of action-oriented cognition: developing process models of the brain's ability to temporarily detach (or disengage) from the here-and-now of the overt sensorimotor loop, as in the case of imagining or planning the future, or mental time travel. These and other detached forms of cognition have been traditionally considered difficult to explain from an action-based perspective and especially in terms of online control loops. In many parts of After Phrenology, Anderson (Reference Anderson2014) suggests that cognition consists in online interaction with the external environment, with no need for neuronal tissue to internalize the structure of such interaction. He also describes aspects of cognition as the manipulation of external symbols (Clark & Chalmers Reference Clark and Chalmers1998) and briefly alludes to the possibility that, if external symbols are unavailable, internal resources such as memory and imagination might be deployed in their stead. This latter possibility, and in general the contribution of internally generated neuronal processing to cognition, would require much more attention to fully understand some aspects of higher cognition as detached cognition.

One domain where internally generated neuronal processing has been studied in great detail is rodent navigation. It has been consistently reported that hippocampal place cells, whose firing is normally associated with the animal's spatial position, can also fire when the animal is outside its standard “place field,” especially during periods of rest or sleep, and at decision points. This “out-of-field” neuronal activity cannot be driven by external stimuli (consider the case of the sleeping animal) but needs to be internally generated based on intrinsic network dynamics. Still, it unfolds in highly organized forms and is neurophysiologically similar to the activity observed during overt action. One example is internally generated sequences of place cell activations observed when the animal rests or sleeps. These sequences can, for example, “replay” spatial trajectories that the animal has experienced (or recombined trajectories), forward or backward, in a time-compressed manner (e.g., during sharp wave ripple complexes, at about 140–200 Hz). These and other forms of internally generated sequences (e.g., preplays) have been associated with various functions such as memory consolidation and planning (Diba & Buzsáki Reference Diba and Buzsáki2007; Dragoi & Tonegawa Reference Dragoi and Tonegawa2011; Gupta et al. Reference Gupta, van der Meer, Touretzky and Redish2010; Pezzulo et al. Reference Pezzulo, van der Meer, Lansink and Pennartz2014; Pfeiffer & Foster Reference Pfeiffer and Foster2013). Another kind of internally generated neuronal sequence is expressed in the theta rhythm (8–12 Hz), when the animals are engaged in decision tasks, and has been associated with “what-if” loops and the anticipation of the consequences of possible choices in order to select, say, one of two arms of a maze (Johnson & Redish Reference Johnson and Redish2007; Wikenheiser & Redish Reference Wikenheiser and Redish2015).

These covert phenomena are not unique to the hippocampus but have been reported in several other brain areas and can at least potentially support a variety of detached or covert cognitive operations (Buzsáki et al. Reference Buzsáki, Peyrache and Kubie2015; Lisman Reference Lisman2015). Importantly, the covert activity recruits the same (or closely related) neurophysiological mechanisms and neuronal resources (e.g., neural assembly sequences) as those implied in overt action, but through internally generated processing. This raises the possibility that detached forms of cognition such as planning, but also potentially others such as imagination and mental time travel, can be explained within an action-based framework rather than require a distinct ontology of neural constructs – if we allow control loops to extend beyond online interactions to also cover internally generated neuronal dynamics.

In this vein, the possibility that the same brain networks can operate and realize control loops in two distinct modes – one stimulus-driven and another (detached) based on internally generated dynamics – has been raised by several researchers and is often associated with the functioning of predictive and generative models (Grush Reference Grush2004; Maye & Engel Reference Maye and Engel2011; Pezzulo & Castelfranchi Reference Pezzulo and Castelfranchi2009) and to theories of reuse (Anderson Reference Anderson2014; Dehaene Reference Dehaene, Dehaene, Duhamel, Hauser and Rizzolatti2005). However, these hypotheses remain to be systematically tested with the help of mechanistic models. If these theories are on the right track, (some) detached forms of cognition might be based on internalized neuronal loops that recapitulate the same dynamics as those learned during overt action (e.g., replays of spatial trajectories in hippocampal ripples) and hence remain fundamentally action-oriented, albeit in an indirect or covert form. The progressive internalization of overt loops, possibly linked to the functioning of internal predictive models, might be one way the neuronal architecture for action control of our ancient evolutionary ancestors progressively developed more sophisticated cognitive abilities in continuity with its more basic sensorimotor skills (e.g., planning and mental time travel abilities on top of the systems that supported spatial navigation, Buzsáki and Moser Reference Buzsáki and Moser2013). This hypothesis would also explain why these apparently disconnected abilities have partially shared neuronal circuits (Schacter et al. Reference Schacter, Addis and Buckner2007; Reference Schacter, Addis, Hassabis, Martin, Spreng and Szpunar2012). Clearly, more specific predictions on the similarities and differences between these abilities are required that can be derived from mechanistic models, which could possibly be tested within the neurophysiological framework for neural coalitions and neural reuse discussed in Anderson (Reference Anderson2014).