Our ability to interact with our environment seems limitless. We can learn to use keyboards—in the office using 10 fingers, at home using only our thumb on the touchscreen of our phone. We can learn to play violin, to drive a car, to do heart surgery, and so on. According to the ideomotor principle of action control, such intention-based actions are performed to produce internally pre-specified and desired effects in the environment (see James, 1890; Stock and Stock, 2004). In this respect, any motor action would result in, or rather from, anticipating its perceptual consequences (Greenwald, 1970; Le Bars et al., 2016).

Pushing the frontiers of natural motor actions, recent advances in neuroscience and engineering are enabling human beings to directly act upon the environment with “thoughts” through Brain-Computer Interfaces (BCI). In a typical non-invasive BCI system, the user’s neural activity is recorded via brain imaging techniques (e.g., EEG, fNIRS, fMRI), before being decoded with computational and Artificial Intelligence (AI) methods. This last phase allows the translation of the brain signals into digital commands that are understandable by the connected device(s) (e.g., a computer, a robot etc.).

Besides the obvious benefit of BCIs for patients suffering from motor impairments, the dramatic expansion of this technology (see Douibi et al., 2021) raises important questions regarding the disembodied nature of resulting actions (Steinert et al., 2019). Notably, one might wonder whether it is even possible to qualify BCI-mediated actions as real human actions, given the potential reduction of sense of agency or responsibility it might cause in users (see Limerick et al., 2014; Rainey et al., 2020). Moreover, it is worth noting that most of non-invasive BCI paradigms aim to enable “acting with thoughts” but do not necessarily respect fundamental aspects of neuroscientific models of human actions, especially regarding the perceptual counterpart of action, which remains barely considered in BCI-mediated actions (see Wang et al., 2019).

In the current article, we attempted to conciliate the neuroscientific models of human actions with non-invasive BCI methods by proposing an innovative and more naturalistic BCI paradigm that would notably take advantage of the ideomotor principle.

In the first section, we summarized the most important evolutions of the ideomotor theory and its alternatives, which all emphasize the importance of action-effect prediction in human action.

In the second section, we reviewed the actual main kinds of non-invasive BCIs and we discussed their limitations in the light of previous motor action models.

Finally, we proposed a new experimental BCI paradigm directly inspired by the ideomotor principle. We believe that (1) adapting BCI paradigms could allow simple action-effect bindings and consequently action-effect predictions and (2) using neural underpinnings of those action-effect predictions as features of interest in Artificial Intelligence (AI) technics, could lead to more accurate and naturalistic BCI-mediated actions.

Recent decades gave rise to a school of thought, which postulates that, in our brain, perceiving our environment and acting upon it is a unified process. Notably, in line with conditioning theories, the ideomotor and common coding principles claim that if actions and their effects are repeatedly contingent, actions would end up in being coded in terms of the effects they evoke in the environment (e.g., Prinz, 1997; Hommel et al., 2001; Waszak et al., 2012).

Such sensorimotor contingencies and action-effect mapping could rely on reinforcement learning (e.g., Colzato et al., 2007; Muhle-Karbe and Krebs, 2012) or active inference (e.g., Friston et al., 2016) processes. Indeed, those models notably emphasize on the fact that action-effect bindings are continuously updated through experience.

In the Ideomotor framework, action is thus conceived of as perceptual states. Not present perceptual states, but future perceptual states. In other words, a voluntary action is supposed to be primarily driven by the anticipated representation of its expected effect or outcome (see Shin et al., 2010 for a comprehensive review).

At a neural level, different theories have linked this action-effect prediction to distinct patterns of brain activations such as the cancelation (see Wolpert et al., 1995), the preactivation (see Waszak et al., 2012), or the sharpening (see Kok et al., 2012) of activity in sensory and perceptual brain areas, which are usually related to the action (see Figure 1).

On the one hand, cancelation of activity in expected sensory areas would result from a forward model in the motor system that allows the prediction of action-outcomes. In fact, this neural cancelation related to perceptual expectations would keep the agent (i.e., the individual performing the action) maximally sensitive to unexpected or important outcomes, to optimize learning or planning of new actions (see Wolpert et al., 1995; Stanley et al., 2007). The cancelation account perfectly matches with the well-known phenomenon of sensory attenuation, which corresponds to a decrease in neural activity (e.g., Baess et al., 2011; Klaffehn et al., 2019) or perception performances (e.g., Cardoso-Leite et al., 2010), and which is commonly associated to the processing of expected outcomes.

On the other hand, the preactivation theory states that action-effect prediction would result from an early enhancement of activity in perceptual areas, which are typically involved in the sensory processing of the action outcome, even though it has not occurred yet in the environment (see Waszak et al., 2012; Roussel et al., 2013, 2014). Previous experiments have successfully demonstrated the existence of neural activation in the sensory units linked to the action, by using brain imaging techniques such as fMRI (e.g., Kühn and Brass, 2010; Kühn et al., 2011) and EEG (e.g., Hughes and Waszak, 2014). Thus, from this point of view, sensory attenuation would result from a more complex discrimination between the observed action-effect and the pre-activation of the predicted effect that occurs before, relative to when no prediction—or misprediction (see Hsu et al., 2015)—is present.

Halfway between the two last principles of cancelation and preactivation, the sharpening account (Kok et al., 2012) argued that sensory expectation would rather result from a suppression of activity in perceptual units that are not supposed to be involved in the processing of forthcoming effects, while neural activity in expected perceptual areas would remain high in parallel. Further experimental evidence of this model was provided recently via fMRI (Yon et al., 2018).

Overall, these neuroscientific theories related to motor action converge toward the two following notions. First, the perceptual component of action is essential and inseparable from motor control, which is notably observable through the action-effect prediction mechanism. Second, the action-effect prediction necessarily relies on the modulation of activity in perceptual brain areas, as synthesized in Figure 1.

However, the precise temporal dynamics of this anticipated action-effect representation is still under debate (see Waszak et al., 2012; Desantis et al., 2014). For instance, some experimental evidence suggested that action-effect representation was activated relatively late, during the action planning step (e.g., Ziessler and Nattkemper, 2011), while other studies—in line with the strong version of the Ideomotor principle—demonstrated that it should occur earlier, during the action selection stage, and before the action initiation (e.g., Paelecke and Kunde, 2007; Le Bars et al., 2016; Dignath et al., 2020).

To our knowledge, no existing BCI paradigm is congruent with the strong Ideomotor view, by using the EEG markers related to the perceptual prediction occurring before the BCI-mediated action.

Three main kinds of non-invasive BCI paradigms are commonly described in literature, depending on the agent’s task and the patterns of brain activity that are consequently generated (see Kögel et al., 2019). First, passive BCIs rely on brain activity that is not voluntarily produced by the user, to monitor his/her neurocognitive or affective state and adequately adapt the environment or issue a warning (e.g., Zander and Krol, 2017). Second, reactive BCIs are based on brain activity changes reflecting the agent’s voluntarily focused attention on a specific external stimulus (e.g., Guger et al., 2012). Third, active BCIs require the user to apply intentionally a particular mental strategy, such as motor imagery (MI) that usually implies imagining a limb movement without actually performing it (MI-BCI; e.g., Salvaris and Haggard, 2014).

Although these three BCI paradigms eventually lead to outputs or new events in the users’ environment, only reactive and active BCI paradigms allow the agents to intentionally perform those changes, by linking a prior intention to act, to a final effect or consequence (Metzinger, 2013; Steinert et al., 2019). Assuming that any voluntary motor action relies on the implementation of the agent’s intention(s) (see Pacherie, 2008), we have thus specifically scrutinized active and reactive BCI-mediated actions through the prism of the motor action models we briefly described in the previous section.

Reactive BCI requires the user to intentionally focus his/her attention on external stimuli to potentially control an external device. In such a reactive paradigm, neural modulations are exogenously generated by specific stimulations provided by the BCI system (e.g., Höhne et al., 2011). Thus, EEG-based BCIs rely on visual or auditory evoked potentials such as the P300 event-related potential (ERP) that occurs 300 ms after an important event (see Jin et al., 2012; Lin et al., 2018), the error-related potential (ErrP) that occurs around 200 ms after a mistake (e.g., Dal Seno et al., 2010; Yousefi et al., 2018) or steady state evoked potentials (SSEP) that are induced by oscillating stimuli (e.g., Chen et al., 2017). Interestingly, some of these reactive settings do rely on neural prediction mechanisms. Notably, P300 and ErrP are, respectively, observed when a rare event (i.e., an event inducing “surprise”) or an erroneous effect (i.e., the actual effect was mispredicted) is presented. For instance, reinforcement learning based BCIs, which use neural error responses (e.g., ErrP signals) as reward feedbacks on the agent’s action, offer the potentiality to get autonomous paradigms that would dynamically adapt in case of erroneous classifications (e.g., Marsh et al., 2015; Wirth et al., 2019).

However, in such reactive BCIs, the used brain potentials do not constitute direct but rather retrospective markers of neural prediction, since they occur after the event or the action execution, as the result of the comparison process occurring between the internal model of forthcoming events (prediction) and the actual sensory effect/event (see Wolpert et al., 1995; Hsu et al., 2015).

Counter to reactive BCI, active BCI is self-paced and allows the user to endogenously/voluntarily control his/her brain activity—and consequently the external connected device (e.g., robot or computer)—at any time, without being tied to a stimulus (see Rao, 2013). In particular, active BCI mainly relies on motor imagery (MI), which can correspond to a visual or a kinesthetic representation of the motor action (see Neuper et al., 2005). Indeed, MI produces neural activity that is spatiotemporally similar to the activity generated during the actual movement even though it is smaller in amplitude (see Pfurtscheller and Neuper, 1997; Miller et al., 2010). Specifically, MI is linked to recognizable EEG brainwave patterns (see Wierzgała et al., 2018) such as a decrease in the frequency bands μ (8–12 Hz) and β (18–30 Hz) corresponding to Sensorimotor Rhythms (SMR). Notably, SMR decrease occurs in the brain hemisphere that is contralateral to the limb “imagined” movement. One might presume that the use of brain mechanisms that are at play in motor execution should be sufficient to ensure a certain theoretical proximity between MI-BCI mediated actions and natural motor actions. However, the representation of perceptual action goal/outcome remains barely considered in such paradigms, and only a few researchers have investigated the importance of implementing action feedbacks to BCIs (e.g., Quick et al., 2020). From a broader perspective, the mental act that is executed through the MI-BCI is frequently disconnected from the final purpose or the proprioceptive effect of that action (see Beursken, 2013; Jeunet et al., 2016). This could notably explain the high illiteracy rate—corresponding to the inability to control the BCI system—that is commonly observed in participants (see Lee et al., 2019).

Thus, on the one hand, many reactive BCIs are based on retrospective markers of sensory prediction occurring after a particular event (e.g., P300 or ErrP). On the other hand, most of active BCIs rely on neural activity associated to sensorimotor commands, without considering the importance of the representation of that motor action in terms of its perceptual consequences.

In this article, we proposed an experimental framework, in line with the ideomotor principle, to test a new type of non-invasive BCI that would allow to perform more naturalistic BCI-mediated actions.

This perspective depends entirely on the ability to decode efficiently the users’ perceptual intentions, by using neural modulations related to action-effect predictions occurring before the action (e.g., Kühn et al., 2010; Waszak et al., 2012; Dignath et al., 2020). This point is not trivial for at least two reasons. First, due to the lack of a unified theory regarding the neural patterns underlying action-effect prediction, which could for instance correspond to a cancelation (Wolpert et al., 1995), a preactivation (Waszak et al., 2012), or a sharpening (Yon et al., 2018) of activity in the perceptual units concerned. Second, because the standard method for brain activity recording in non-invasive BCIs, namely EEG, has a low spatial resolution, which prevents from a precise discrimination between different sources of activations in perceptual units. Thus, further studies must harness adequate paradigms to allow perceptual intentions modeling, notably by using sensory feedbacks that would elicit discriminable EEG patterns during action-effect prediction and that would also be discriminable from brain activations resulting from the BCI task itself.

Theoretically, an independent ideomotor BCI would then constitute a novel BCI paradigm located in-between active, reactive and passive paradigms, being endogenously controlled as active paradigms, i.e., the agent would have to think about the perceptual feedback of the BCI-mediated action, but relying on sensory stimuli, i.e., the action-effects, as reactive paradigms. One might also propose to use this ideomotor paradigm in combination with other BCI methods (e.g., MI-BCI). Given the automatic nature of the action-effect prediction process (see Kunde, 2004; Le Bars et al., 2016), the agent would not be instructed to make a conscious representation of the intended feedback, even though EEG markers of action-effect prediction would be used as additional features to improve the BCI accuracy. In this case, the hybrid ideomotor BCI would be closer from passive BCI paradigms but would still rely on sensory stimuli, as represented in Figure 3.

In conclusion, an ideomotor BCI would allow to get closer from natural action by considering the perceptual facet of human action through action-effect prediction. This should improve both users’ experience and systems accuracy. Importantly, an effective BCI paradigm—that would be based on perceptual representation of action would definitely endorse and concomitantly renew the old-fashioned ideomotor theory.

The original contributions presented in the study are included in the article/supplementary material, further inquiries can be directed to the corresponding author/s.

SL and FW developed the idea and the research proposal which were described in the current article. SL wrote the manuscript with supervision from SC and FW. All authors contributed to the article and approved the submitted version.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.