In Vigor: Neuroeconomics of Movement Control, Shadmehr and Ahmed convincingly argue that movement vigor can be used as a proxy to measure how the brain assigns value to action goals (Shadmehr & Ahmed, Reference Shadmehr and Ahmed2020). The authors demonstrate that because goal-oriented decisions and actions involve some kind of “economic” computations in which effort, time, and reward are evaluated, subjects react and move more vigorously when the option is more desirable (i.e., they value it more), and conversely slow their behavior in the prospect of unpleasant events.

Although appealing, this proposal lacks an important precision in my opinion: What is vigor, and how to measure it in the first place? The notion of vigor is rather intuitive, but it's commonly associated with very different behavioral variables in the system neuroscience literature. Vigor sometimes refers to movement speed and/or duration (Muhammed, Dalmaijer, Manohar, & Husain, Reference Muhammed, Dalmaijer, Manohar and Husain2020; Thura, Cos, Trung, & Cisek, Reference Thura, Cos, Trung and Cisek2014), to speed or duration scaled by amplitude (Baraduc, Thobois, Gan, Broussolle, & Desmurget, Reference Baraduc, Thobois, Gan, Broussolle and Desmurget2013; Berret, Castanier, Bastide, & Deroche, Reference Berret, Castanier, Bastide and Deroche2018; Choi, Vaswani, & Shadmehr, Reference Choi, Vaswani and Shadmehr2014; Reppert, Lempert, Glimcher, & Shadmehr, Reference Reppert, Lempert, Glimcher and Shadmehr2015; Reppert et al., Reference Reppert, Rigas, Herzfeld, Sedaghat-Nejad, Komogortsev and Shadmehr2018; Thura, Reference Thura2020), to movement kinematics and reaction times (Milstein & Dorris, Reference Milstein and Dorris2007; Sedaghat-Nejad, Herzfeld, & Shadmehr, Reference Sedaghat-Nejad, Herzfeld and Shadmehr2019; Summerside, Shadmehr, & Ahmed, Reference Summerside, Shadmehr and Ahmed2018), or to reaction times only (Griffiths & Beierholm, Reference Griffiths and Beierholm2017; Guitart-Masip, Beierholm, Dolan, Duzel, & Dayan, Reference Guitart-Masip, Beierholm, Dolan, Duzel and Dayan2011). Vigor is also sometimes qualified via the rate of responses (Dezfouli, Balleine, & Nock, Reference Dezfouli, Balleine and Nock2019; Niv, Daw, Joel, & Dayan, Reference Niv, Daw, Joel and Dayan2007), the level of engagement, arousal, or motivation in a task (Mazzoni, Hristova, & Krakauer, Reference Mazzoni, Hristova and Krakauer2007; Turner & Desmurget, Reference Turner and Desmurget2010).

Despite a lack of consistent definition, it is safe, however, to say that most of the vigor-related metrics listed above strongly co-vary with each other: Increasing movement speed for a given amplitude typically reduces its duration and increases its energetic cost; or increasing movement amplitude increases its duration and speed. The same problem arises for the notion of “value,” which is naturally confounded by many factors, including pleasure, curiosity, pain, and so on, and for which the existence of a proper neural substrate is still debated despite decades of investigations (Ballesta, Shi, Conen, & Padoa-Schioppa, Reference Ballesta, Shi, Conen and Padoa-Schioppa2020; Hayden & Niv, Reference Hayden and Niv2020). In their book, Shadmehr and Ahmed demonstrate that movement vigor is actually another confounding factor that one should take into account in order to identify the neural circuits of valuation. Recent publications even suggest that valuation and motor invigoration are two undistinguishable functions, mediated by shared/common neural networks (Collins & Frank, Reference Collins and Frank2016; Hayden & Niv, Reference Hayden and Niv2020).

Beyond semantic debates (Baker, Lansdell, & Kording, Reference Baker, Lansdell and Kording2021; Brette, Reference Brette2019; Cisek, Reference Cisek1999), one might wonder why the brain would bother having to compute multiple discrete quantities that are all so tightly linked? Wouldn't it benefit, in terms of accuracy-energy expenditure trade-off, from computing fewer but integrated quantities that capture, if not all, most of parameters variability? In this view, movement vigor would approximate the reciprocal of the chosen behavior duration (or its speed) that integrates the context-dependent reward rate computation based on costs (opportunity and energy) and rewards. Because each individual perceives reward and costs differently, vigor is idiosyncratic in nature (Berret et al., Reference Berret, Castanier, Bastide and Deroche2018; Reppert et al., Reference Reppert, Rigas, Herzfeld, Sedaghat-Nejad, Komogortsev and Shadmehr2018). And because in a given task, different motor repertoires serve different goals (e.g., exploration vs. interaction), oculomotor vigor should not necessarily co-vary with arm movement vigor, or locomotion vigor. Thura et al. (Reference Thura, Cos, Trung and Cisek2014), for instance, report in monkey that saccade velocity is less influenced by reward rate than reach velocity in a reaching decision-task. Reppert et al.'s (Reference Reppert, Rigas, Herzfeld, Sedaghat-Nejad, Komogortsev and Shadmehr2018) also show in human that vigor, defined here as a “latent” variable, is shared among skeletomuscular motor modalities, but has no relationship with oculomotor behavior. Thus, by establishing a desired level of vigor for a given goal, the brain does not necessarily need to separately compute movement planning and execution durations, speed, and acceleration, and can even maybe flexibly exchange deliberation time for movement planning/execution time depending on the subject's needs and preferences (Reynaud, Saleri Lunazzi, & Thura, Reference Reynaud, Saleri Lunazzi and Thura2020).

Most of the recent neurophysiological results fit an integrated view of neural organization (Cisek & Kalaska, Reference Cisek and Kalaska2010). For example, studies on monkey have long reported that dorsal premotor (PMd) and primary motor (M1) neurons vary their activity as a function of movement duration, velocity, acceleration and distance (see Kalaska, Reference Kalaska and Sternad2009 for a review). More surprisingly, these motor neurons also “encode” estimation of elapsing time (Renoult, Roux, & Riehle, Reference Renoult, Roux and Riehle2006) and reward (Ramkumar, Dekleva, Cooler, Miller, & Kording, Reference Ramkumar, Dekleva, Cooler, Miller and Kording2016). Perhaps even more dramatically, dopamine neurons have been shown to be crucial for action valuation during reinforcement learning (e.g., Glimcher, Dorris, & Bayer, Reference Glimcher, Dorris and Bayer2005; Lee, Seo, & Jung, Reference Lee, Seo and Jung2012), to be a key component of the neural circuits that regulate the effort and motivation required to obtain a reward (e.g., Jurado-Parras et al., Reference Jurado-Parras, Safaie, Sarno, Louis, Karoutchi, Berret and Robbe2020; Salamone et al., Reference Salamone, Pardo, Yohn, López-Cruz, SanMiguel, Correa, Simpson and Balsam2016), and to determine the level of vigor with which the selected movement is executed (Collins & Frank, Reference Collins and Frank2016; da Silva, Tecuapetla, Paixão, & Costa, Reference da Silva, Tecuapetla, Paixão and Costa2018; Panigrahi et al., Reference Panigrahi, Martin, Li, Graves, Vollmer, Olson and Dudman2015; Rueda-Orozco & Robbe, Reference Rueda-Orozco and Robbe2015; Yttri & Dudman, Reference Yttri and Dudman2016; Zenon, Devesse, & Olivier, Reference Zenon, Devesse and Olivier2016).

How such a degree of neural integration, at least for some brain areas, is related to “good-enough” policies at the behavior level is an open question. Yet, monkey and human studies indicate that subjects rarely aim for “perfection” or optimality during tasks execution in the lab. In the motor domain for instance, participants typically perform highly variable movements (in terms of speed and accuracy) as long as those actions still lead to rewards. Similarly, human decision-makers easily adopt heuristics approaches in a large number of decision conditions, sometimes more successfully than if they opted for a rational strategy (Gigerenzer & Gaissmaier, Reference Gigerenzer and Gaissmaier2011). In the optimal foraging context, Shadmehr and Ahmed (Reference Shadmehr and Ahmed2020) note that “the actual behavior of animal disagrees with the exact formulation of the theory in a number of experiments” (p. 73). Therefore, many situations seem to favor goal-driven good-enough policies dictated by an accuracy-energy expenditure trade-off much more than rational and normative operations requiring a vast a priori knowledge of the world as well as costly computations of multiple low-level behavioral parameters. Interestingly, some sub-optimal behavioral patterns have been proposed to emerge from pre-wired neural organizations and pre-defined computations (Buzsáki, Reference Buzsáki2019).

A given behavior emerges because integrated, distributed, and recurrent neural activities take place in a body surrounded by an environment, in service of a goal (Gomez-Marin & Ghazanfar, Reference Gomez-Marin and Ghazanfar2019). We thus need to address behavior from the viewpoint of the brain, considering why and how its species-specific circuits successfully evolved over time to control interactions with the world (Cisek, Reference Cisek2019; Jourjine & Hoekstra, Reference Jourjine and Hoekstra2021). To this end, neglecting behavior would obviously be an error (Cisek & Kalaska, Reference Cisek and Kalaska2010; Krakauer, Ghazanfar, Gomez-Marin, MacIver, & Poeppel, Reference Krakauer, Ghazanfar, Gomez-Marin, MacIver and Poeppel2017; Niv, Reference Niv2020). But could it be that “over-analyzing” behavior turns out to be ultimately more misleading than insightful? We all know that studying behavior in a lab means immediately giving up many aspects of real-life behavior. In addition, because behaviors are complex and variable, tasks often decouple parameters in a given number of dimensions of the full outputs space, allowing one to observe whether neural activity follows more closely one or the other of the decoupled parameters (Reimer & Hatsopoulos, Reference Reimer and Hatsopoulos2009). In such tasks though, animal's behavior deviates even further from its ethological repertoire, calling into question our ability to ultimately map this behavior with the neural circuits shaped by millions years of interactions with the external world in natural conditions. Even when a task is ethologically compatible, one should avoid, at least initially, to over-decompose behavior during data analysis, considering that the same neural circuit can generate different behavior patterns as long as the animal's goal is reached (Katz, Reference Katz2016). In other words, just as dimension reduction of neural activities may reveal how neural circuits operate and compute (Humphries, Reference Humphries2020; Kalaska, Reference Kalaska2019; Saxena & Cunningham, Reference Saxena and Cunningham2019; Vyas, Golub, Sussillo, & Shenoy, Reference Vyas, Golub, Sussillo and Shenoy2020), a heuristics-oriented “dimensions reduction” of the behavioral variables under investigation might be fruitful too if one hopes to eventually shed light on the brain–behavior mapping mystery.

One way to implement this strategy in neuroscience studies is to complement and compare “controlled” experiments with tests made in “naturalistic” settings, using hypothesis-based and ethologically-relevant tasks, and to initially analyze integrated behavioral parameters only. Although more challenging technically, especially for studies on large animals, recent developments in unconstrained in-vivo electrophysiology methods (Berger, Agha, & Gail, Reference Berger, Agha and Gail2020) combined with virtual reality systems (Dombeck & Reiser, Reference Dombeck and Reiser2012; Noel et al., Reference Noel, Caziot, Bruni, Fitzgerald, Avila and Angelaki2021) or powerful AI-based pose estimation algorithms (Mathis et al., Reference Mathis, Mamidanna, Cury, Abe, Murthy, Mathis and Bethge2018) should make this naturalistic approach more accessible in the near future.

For example, a naturalistic setting would allow one to test a recent hypothesis about the role of the monkey basal ganglia (BG) in the selective invigoration of behavioral repertoires (Cisek & Thura, Reference Cisek, Thura, Corbetta and Santello2018). In this proposal, the monkey premotor/motor and posterior parietal cortices are organized in “zones” of ethologically relevant types of motor repertoires, such as reaching or climbing movements (Graziano, Reference Graziano2006). The dopamine-dependent putamen neurons receive massive inputs from all cortical areas and project to downstream structures of the BG organized as segregated functional modules (Grillner & Robertson, Reference Grillner and Robertson2016) connected to a given cortical zone, via connections in the thalamus. Now imagine a cage designed so that a monkey can adopt a “foraging” behavior and interact at its convenience with two touch screens, one accessible from the lower part of the cage, the other placed higher up. Each screen displays visual stimuli associated with evolving cost/reward contingencies. The monkey would have the choice to directly reach the screen in front of him (“exploitation” decisions, within a given behavioral repertoire) or, as the animal would do in a tree, move in the cage (“exploration” decisions, between behavioral repertoires) to place himself in front of the other screen and start a new sequence of exploitation decisions. With wireless neural recordings and a set of video cameras installed in the cage, we could assess whether during exploitation, the reaching repertoire is selected and invigorated whereas the other repertoires (climbing, walking, eating, etc.) are suppressed through the cortico-BG-thalamo-cortical loops, and whether deliberations within the selected repertoire are resolved through competitions in cortical areas (Cisek, Reference Cisek2007; Thura & Cisek, Reference Thura and Cisek2017).

To summarize, I argue in this commentary that behavioral neuroscience, including my own work, suffers from a temptation to “disintegrate” behavior so that artificially defined subsystems appear more easy to investigate. I believe that such an approach confuses terminology, encourages unnatural behavior, and eventually limits the production of insightful evidence about what neural systems have been designed for. I agree with Shadmehr and Ahmed that movement vigor informs us about how we value our actions. However, as an evaluation tool, a clear definition of vigor and quantification methods is lacking. I believe this is because vigor is one of the low-dimensional parameters that approximate how long and/or how fast the brain interacts with its environment given particular circumstances. Consequently, vigor cannot be precisely quantified, but only roughly characterized based on arbitrary-selected decision and/or movement variables.