The functional role of conscious sensation of movement

This paper proposes a new framework for investigating neural signals sufficient for a conscious sensation of movement and their role in motor control. We focus on signals sufficient for proprioceptive awareness, particularly from muscle spindle activation and from primary motor cortex (M1). Our review of muscle vibration studies reveals that afferent signals alone can induce conscious sensations of movement. Similarly, studies employing peripheral nerve blocks suggest that efferent signals from M1 are sufficient for sensations of movement. On this basis, we show that competing theories of motor control assign different roles to sensation of movement. According to motor command theories, sensation of movement corresponds to an estimation of the current state based on afferent signals, efferent signals, and predictions. In contrast, within active inference architectures, sensations correspond to proprioceptive predictions driven by efferent signals from M1. The focus on sensation of movement provides a way to critically compare and evaluate the two theories. Our analysis offers new insights into the functional roles of movement sensations in motor control and consciousness.


Follow the signals: studying conscious proprioceptive sensations and their role in motor control
When a person moves their body to grasp something, throw something, or maybe kick something, it feels like something.Moving one's body gives rise to bodily sensations.Our goal with this article is to provide a new framework for studying the role of conscious sensation of movement in the control of movements.We propose a strategy of focusing on the type of sensory information and physiological processes that can be sufficient for a conscious sensation of movement.On this basis, we locate the most likely functional roles in the two major theories of motor control in contemporary motor control research, namely, motor command theories (such as optimal control theory) and active inference theories.The two types of motor control theory allocate different locations to conscious sensations of movement in their functional architectures and thus provide different explanations.We show that a focus on conscious sensation of movement has important implications for current discussions of motor control and theories of consciousness.
Losing one's ability to sense one's bodily movements dramatically alters one's ability for action (Wong, 2018).Typically, patients who have suffered complete proprioceptive deafferentation but have retained intact motor circuitry will at first lose their ability to perform voluntary movements.Patients cannot control the extent and vigor of movement.They cannot focus on and activate individual muscle groups.With substantial self-discipline and effort, some patients might reacquire a capacity for voluntary bodily action, but the type of ability will have changed.Patients will be able to move only when movements are visually guided and with immense mental concentration and explicit planning.Sneezing will be disastrous; walking while daydreaming will be impossible (Cole, 1995).Movements will have lost their precision and fluency.In the words of Ian Waterman: "I always have to think where I am and what I'm doing.I can never be completely absorbed.However beautiful the scene I have to concentrate on what I'm doing […].You might remember a walk for the views.I just remember the walking" (Cole, 1995, 129).This indicates that conscious proprioceptive sensations of movement accompany our active life in the background and play an important role in the control of bodily action.Which role conscious proprioceptive sensations are playing in motor control and how best to describe and model the role remains an enigma.
There are at least two important reasons for the enigmatic role of conscious sensation of movement in control of movement.First, despite important results and progress of contemporary science of consciousness (Seth and Bayne, 2022;Fazekas et al., 2023;Northoff and Lamme, 2020), research has been almost exclusively focused on visual consciousness.Almost all contemporary theories and controversies concern visual consciousness.Important issues like the neural location, the role of attention, and processing latencies are framed in terms of visual processing.The field of consciousness research has been characterized by an almost total neglect of conscious proprioception.
Second, research on motor control has typically been conducted with a complete lack of interest in the conscious aspect of proprioceptive sensation.Many aspects of movement control can be performed in the absence of conscious awareness.Human motor neuroscience has a good understanding of the types of information that enable eye-hand coordination in successful reaching and grasping (Bekkering and Sailer, 2002;Weiler and Pruszynski, 2020).Reaching and grasping movements are subject to natural variability.When unforeseen internal or external forces perturb the movement, fast online corrections are performed (Crevecoeur et al., 2012;Pruszynski and Scott, 2012).Consequently, often, the last homing phase of the movement requires that retinal information about the location of the target is integrated with proprioceptive information from muscles (Lavrysen et al., 2018;Starkes et al., 2002).These forms of online adjustment require that the central nervous system (CNS) can use proprioceptive information from the muscles to estimate the orientation of the limb, as well as its movement direction, velocity, and force.Typically, researchers have assumed that these control processes and the proprioceptive information are nonconscious (Cordo et al., 1995).In this paper, we will focus on the issue of motor control.
Experimental results, including results from our own group (Brandt et al., 2024), are changing this situation.The results can tell us what kind of information and physiological processes are sufficient for a conscious sensation of movement.In this paper, we focus on two sets of experiments and one plausible implication.The first set of experiments, often using muscle vibration, demonstrates that afferent signals from muscle spindles are sufficient to create clear and robust sensations of limb movement with specific direction, force, and velocity.The second set of experiments, often using momentary deafferentation by peripheral nerve blocks, demonstrates that efferent motor signals from primary motor cortex (M1) are sufficient to create similar conscious sensations of movement, even in the absence of proprioceptive feedback signals.We review the relevant experimental literature.On this basis, we show that a plausible implication of the two sets of experimental studies is that active sensation of movement (involved in voluntary movement) is phenomenologically and physiologically distinct from passive sensation of movement (involved in passive limb displacement).As we will see later, several studies provide independent support for the claim that active and passive movements are associated with distinct types of sensation (Mima et al., 1999;Paillard and Brouchon, 1968).This gives us two important facts ("data points") about sensation of movement and one plausible implication: afferent signals from muscle spindles are sufficient for conscious sensation of movement, efferent motor signals from M1 are sufficient for similar conscious sensations, and active and passive sensation of movement are distinct.
By adopting this strategy of following the sensory signals, we can describe a role for conscious sensations of movement in the control of movements like reaching and grasping.Focusing on results from experiments with sensation of movement allows us to locate likely places for the sensations of movement in the functional architectures of two major competing theories of motor control, viz. the motor command theory and the active inference theory.This enables us to ask how the theories would account for the experimental results (the two "data points") and the implication that active and passive movements are associated with distinct types of sensation.Our goal is to show that a focus on results from experiments on sensation of movement could be a decisive factor in our evaluation and choice of theory of motor control.

Signals in sensation of movement
Our strategy is to explain the sensation of movement by the involved sensorimotor signals.We can explain the contribution of a particular source of information or signal to the sensation of movement by manipulating the signal (for instance, the sensory signals from muscle spindles) while holding other types of involved signals (for instance, afferent visual signals or efferent motor signals) constant or omitting them.In this section, we follow the signals.

First data point: afferent signals and the conscious sensation of movement
Proprioceptive signals about the position of one's limb as well the direction of its movement, velocity, and force play crucial roles in reaching and grasping (Weiler and Pruszynski, 2020).In a noisy and dynamic environment, successful reaching and grasping require the ability to quickly adjust the movements and posture in the face of perturbations.Proprioceptive signals are crucial for this kind of adjustment.The signals not only provide the brain with information enabling it to quickly estimate the current state of a bodily effector, but it also enables it to predict future states of the effector (Pruszynski and Scott, 2012).
What exactly are the proprioceptive signals that are relevant for control of reaching and grasping?When muscles contract, nerve endings of muscle spindle fibres in the intrafusal muscle fibres transform the mechanical stretch of the muscle into electrical signals.When the stretching of muscles increases, the muscle spindles increase their firing rate (Matthews, 1964).When load is exerted on a muscle, Golgi tendon organs in the muscle-tendon junction become squeezed by collagen fibres, and the Golgi tendon organs will increase their firing rate (Matthews, 1933).Joint receptors, located in joint capsules and in the ligaments and terminal regions of tendons, are a third class of proprioceptive receptors (Andrew and Dodt, 1953;Tracey, 1978).In addition, contraction of muscles will also give rise to signals from somatosensory receptors in the skin.These are located both superficially and in deep skin tissue and are mechanically perturbed by the movement of a limb (Burke et al., 1988;Hulliger et al., 1979).This means that a muscle contraction leads not only to the movement of a limb but also to the sensory feedback from receptors in muscle spindles, tendons, joints, and skin.For convenience, we use the term "proprioceptive signals" to refer to afferent signals from all four groups of receptors.By combining the information from the proprioceptive receptors, it is possible for the CNS to derive approximate representation of limb position, limb movement, movement direction, movement velocity, and force (Proske and Gandevia, 2012;Verschueren et al., 1998).It remains a difficult computational puzzle exactly how signals from the various proprioceptors are combined to yield these representations (Crevecoeur et al., 2016;Kavounoudias, 2017).
Motor physiologists generally distinguish between the proprioceptive sense of position and the proprioceptive sense of movement (Blanchard et al., 2011;Houk et al., 1981;Verschueren et al., 1998).This distinction is supported by experimental studies where participants are unable to detect movements of their limbs while still being able to detect changes in position (Refshauge et al., 1995).Furthermore, experimental studies show that we can separately induce illusions of movement and illusions of position by manipulating the amplitude and frequency of vibration on a given muscle (McCloskey, 1973).In the present paper, we are mainly interested in the sense of movement.
In the scientific discussions of bodily movement, many physiologists accept that humans experience bodily sensations when they move.For instance, Goodwin and colleagues write: Under most conditions we are consciously aware of the position of the various parts of our limbs relative to each other and whether they are moving or still.This awareness has been given, among others, the names of "kinaesthesia" and "position sense."(Goodwin et al., 1972a, p. 705, we added the italics) Or take a more recent review by Naito: The somatic sensation of limb movement is normally experienced when a limb is actually moved.This is because the sensory afferents from the muscle spindles, cutaneous receptors, and joint receptors increase their activities during passive and active limb movements, and the afferents convey the somatic information of limb movements to the brain.In particular, signals from the muscle spindles play very important roles in kinesthesia, the perception of limb movements.(Naito, 2004, p. 73, we added the emphasis and removed all references in the original) In short, when your hand is moved from the computer keyboard to the monitor, you have conscious sensations of movement in your hand and arm primarily due to the activation of sensors in the muscle spindles, the tendons, the joints, and the skin of your wrist and arm.
One standard way to study the content of the information conveyed from proprioceptive receptors, in particular from receptors in the muscle spindles, to the cortex is by inducing sensory illusions of movement (see Figs. 1B and 2B).By mechanically vibrating muscles, researchers can activate the muscle spindles and produce sensations of movement (Goodwin et al., 1972b;McCloskey, 1973;Roll and Vedel, 1982;Tidoni et al., 2015).For instance, by vibrating the biceps with the right range of frequencies, the vibration stimulates the muscle spindles in ways similar to a situation where the muscle is extending.The result can be a strong sensory illusion of arm extension, even when the arm is immobile and the participant is aware of this fact (Albert et al., 2006;Roll et al., 1989).When vibrations are applied to several appropriately chosen muscle groups around the wrist, it can produce the sensory illusions of complex movements such as drawing or writing movements (Roll et al., 2009).
The sensations of movement are usually studied by vibrating the appropriate muscles and asking the participant to report the position of the stimulated limb.A standard set-up would be one in which a blindfolded participant is asked to keep her right arm in a position parallel to the sensed position and movement of her left arm, which is restrained and then stimulated to produce a sensation of movement.Researchers generally assume that the sensations elicited in vibration experiments have a representational content and are illusory counterparts of normal veridical sensations of movement.The representational content of the sensations of movement is sufficiently fine-grained to allow the person to accurately match their unstimulated limb to the sensed position, movement direction, and velocity of the stimulated limb.
Summing up, the experimental approaches discussed in this section are often assumed to study the content of the information sent from proprioceptive receptors to the cortex, giving rise to the conscious Fig. 1.Motor command theory: General architecture.A shows signal processing in an intact central nervous system during a voluntary movement.The orange ellipse indicates where in the network a signal estimates the current state of the body which can give rise to a sensation of movement.In the intact central nervous system, the ellipse encompasses corollary discharge signals from the forward model, sensory feedback signals from the periphery, and error signals after comparing predicted and actual proprioceptive feedback signals computed in the primary somatosensory cortex (S1).B shows the same network under muscle tendon vibration.The same location indicated by the orange ellipse may correspond to the illusory sensation of movement induced by muscle tendon vibration (e.g., Roll and Vedel, 1982).C shows the same network undergoing TMS stimulation during peripheral nerve block (e.g., Amassian et al., 1989;Christensen et al., 2010) again with the orange ellipse indicating a location that may correspond to a sensation of movement.D shows the same network during an attempt to make a voluntary movement with peripheral nerve block (e.g.Gandevia et al., 2006).Again, the same location of the network corresponds to a sensation of movement.E shows the same network in a deafferented person.In this network, it is unclear from standard versions of the motor command theory what the forward model may signal, hence the question mark next to 'corollary discharge'.PFC: Prefrontal cortex.PC: Parietal cortex.S1: Primary somatosensory cortex.M1: Primary motor cortex.PMC: Premotor cortex.Cb: Cerebellum.
T. Grünbaum and M.S. Christensen sensations of movement that accompany most, if not all, normal movements of the body.The important point to notice is that the afferent proprioceptive signals are sufficient to generate a robust sensation of movement.Even in the absence of actual movement of the limb, the signals are sufficient to generate a conscious sensation of the limb orientation as well as movement direction, velocity, and force.This is our first data point.

Second data point: efferent signals and the conscious sensation of movement
Passive movements of arms and hands activate proprioceptive receptors in muscles, joints, ligaments, and skin, and convey proprioceptive feedback signals to the somatosensory systems of the brain, eventually giving rise to conscious sensations of movement.Active movements are more complex.Many organisms must be able to distinguish between passive displacement and active movement of its limbs.It has long been thought that efferent motor command signals play an important computational role in distinguishing between other-caused and self-caused limb movements.The role of efferent motor signals in the sensation of movement has been far more controversial.
One central question is whether efferent motor signals are sufficient for a sensation of movement or only for a sense of effort.The sense of force has traditionally been identified with afferent signals, in particular signals from receptors in Golgi tendon organs (Jones, 2003).By contrast, the sense of effort has been associated with efferent signals (Gandevia, 1996).For a long time, consensus in the scientific literature was that efferent signals can only give rise to sense of effort, though some researchers suggested that the signals might be sufficient to produce a sensation of position or displacement.For instance, Merton (1964) studied the sensations of thumb movement.The participants had their joint and skin of the thumb anaesthetised with an anoxic nerve block leaving muscles around the joint intact.According to Merton, the participants experienced sensations of active thumb movement and believed that they had succeeded in moving it, even when the experimenter had externally blocked the movement.In various papers, McCloskey argued that the sensations reported by participants in Merton's study might be explained by afferent signals from non-anaesthetised parts of the body.When controlling for this confound, the effects turned out to be very hard to replicate (Goodwin et al., 1972a;McCloskey and Torda, 1975;McCloskey, 1981).Whereas Merton anaesthetised the participants but left the motor system intact, Gandevia et al. (1993) induced complete paralysis in their participants by injecting them with atracurium.In one set of trials, participants were asked to exert maximal handgrip effort.To the surprise of the authors, in the state of complete paralysis, this exertion produced "clear sensations of movement" in addition to sensations of effort.In a second set of trials, the motor cortex of the participants was stimulated by TMS, which had Fig. 2. Active inference theory: General architecture based on Adams et al., (2013).The orange ellipses indicate where in the network a signal is likely to correspond to a conscious sensation of movement might be located.A question mark indicates that standard versions of active inference theory do not allow us to infer a clear theoretical prediction.A shows signal processing in an intact central nervous system during a voluntary movement.B shows the same network under muscle tendon vibration (e.g., Roll and Vedel, 1982).C shows the same network undergoing TMS stimulation during peripheral nerve block (e.g., Amassian et al., 1989;Christensen et al., 2010).D shows the same network during an attempt to make a voluntary movement with peripheral nerve block (e.g., Gandevia et al., 2006).E shows the same network in a deafferented person.PFC: Prefrontal cortex.PC: Parietal cortex.S1: Primary somatosensory cortex.M1: Primary motor cortex.PMC: Premotor cortex.
the effect of inducing illusory finger twitches.Importantly, the illusory twitches were not accompanied by any sense of effort.It is important to note that in their 1993 study, Gandevia et al. (1993) did not anaesthetise the participants.A possible explanation of the sensation of movement is that the paralysis induced by atracurium did not block the fusimotor system projecting along intrafusal fibres.One explanation is therefore that the sensation of movement is produced by afferent signals from muscle spindles activated by the fusimotor system.
Clear support for the claim that efferent motor signals are sufficient for a conscious sensation of movement requires experimenters to control for all these confounding factors.More recently, Gandevia et al. (2006) were able to induce sensations of wrist movement in participants when their arm was temporarily paralysed and anaesthetised by ischaemia.In their procedure, the arm of the participant was completely anaesthetised and paralysed using a pressure cuff positioned on the upper arm (to ensure complete loss of sensation, two participants received an additional injection of lignocaine).In absence of visual feedback, the task was to exert force in the direction of either flexion or extension of the wrist of their anaesthetised and paralysed hand, and then match the position with their normal hand.According to the authors, they were able to provide "direct evidence that signals related to the motor command or effort can produce an illusion that a part of the body has been displaced even though it has not moved" (708).
The study by Gandevia et al. (2006) carefully controlled for the confounds that confused earlier studies.The study therefore provides convincing evidence for the sufficiency of efferent motor signals to produce a conscious sensation of movement, or at least a sense of position.It is worth noting that this result is consistent with two competing explanations.On the one hand, the efferent driven sensation of movement could be directly related to the efferent motor output.On the other hand, the efferent driven sensation of movement could be the output of a forward modelling process in which the efferent signal is translated into a sensory prediction.According to the latter proposal, the output of this translation process would involve activation of somatosensory areas of cortex (Christensen et al., 2007).Both of these explanations assume a motor command or internal model architecture of motor control.For the various architectures and explanations of Gandevia et al.'s experimental situation, see Figs. 1D and 2D.
Another experimental approach has used non-invasive brain stimulation to induce illusions of movement by applying trains of transcranial magnetic stimulation (TMS) pulses over cortical brain areas involved in generating motor signals (e.g.M1, PMd) (Amassian et al., 1989;Christensen et al., 2010).In the absence of both outgoing motor signals innervating muscles in the forearm and afferent signals providing proprioceptive feedback to the CNS due to temporary ischemic nerve block (caused by an inflated pressure cuff that induces a temporary state of anaesthesia and paralysis), researchers were able to produce TMS induced illusions of movements (participants scored their sensation of movement on a Likert-like scale, Christensen et al., 2010).These studies provide additional support for the claim that centrally generated efferent signals are sufficient to generate illusions of movements.This experimental situation is described in Figs.1C and 2C.This is our second data point.

The choice point: passive vs active movements and the conscious sensation of movement
We have surveyed studies showing that afferent proprioceptive feedback signals and efferent motor signals can each be sufficient for conscious sensation of movement.In normal active movement, afferent and efferent signals are integrated and give rise to a unified proprioceptive sensation of movement.An interesting question is whether the sensations of movement for passive and active movements have a common or a separate type of content.Does the contribution of efferent signals to the sensation of movement in the case of active movement make these sensations different from the case of passive movement?This question has attracted some attention of philosophers (Grünbaum, 2008;Mylopoulos, 2017;Pickard, 2004;Shepherd, 2017).Following the philosophical literature, we can distinguish between a common content view and a separate content view for sensation of movement.Whether a theory implies a common content view or separate content view depends on its basic assumptions about sensorimotor processing.The choice between common content or separate content views can therefore be used to distinguish between the theories and their implications.It is an important choice point.
According to the common content view, active and passive movements are characterized by common type of sensation of movement.Passive and active movement sensation are alike with respect to type of sensory signals and phenomenology.Nothing in the content of the sensations marks the movement as being an active or a passive movement.There are two versions of the common content view.According to the passive version, efferent signals give rise to corollary discharge signals used to predict reafferent consequences of the movement.These predictions are used to attenuate or subtract the efferent contribution.Consequently, only the passive exafferent contribution arises to the level of conscious sensation of movement (Blakemore et al., 1999;Williams and Chapman, 2002).That is, only the non-predicted afferent feedback from muscles produces a sensation of movement, and the efferent signals play no distinctive role in shaping the sensations of movement.If active and passive movements have the same kinematic pattern, they would be associated with the same sensations of movement.According to the active version of the common content view, conscious proprioceptive sensation is a form of proprioceptive prediction and proprioceptive prediction is the function of signals from M1.These signals are operative in both passive and active sensation of movement.So, according to this view, all sensation of movement is sensation of active movement.As we will see below, this might be a consequence of the active inference theory (compare Figs. 2A and B).
According to the separate content view, active movements are associated with a different type of sensation of movement than passive movements.The separate content view is made probable by the results showing that efferent signals might be sufficient for generating a sensation of movement.The results suggest that participants can experience a sensation of movement in the absence of any afferent sensory feedback from muscles, joints, and skin.Overall, an integrative view according to which the efferent signals modulate and significantly shape the integrated content of sensation of active movement is difficult to accommodate by a common content view.
The separate content view receives independent support from an experiment by Christensen and Grünbaum (2018).In this experiment, participants had to judge whether their finger movements were passive or active in a situation where passive and active movements were kinematically matched and both passive and active movements were planned.The results showed that participants very rarely confused passive movements for active movements, despite their apparent kinematic and cognitive confusability.The authors interpreted the results as demonstrating that participants had some access to their motor signals.
Plausibly, this means that active and passive movements feel different.This conclusion receives additional support from numerous studies indicating that active and passive movements are associated with dedicated or dissociable mechanisms for sensory processing (Bays et al., 2005;Christensen et al., 2007;Dubynin et al., 2021;Mima et al., 1999;Paillard and Brouchon, 1968).
Summing up the argument thus far, we have provided arguments for three substantial points.First, we have provided evidence for the data point that afferent proprioceptive signals are sufficient for conscious sensations of movement.Second, we have provided evidence for the data point that efferent motor signals are sufficient for conscious sensations of movements or at least a sensation of position.Third, we have pointed to the choice point forcing us to choose a common content or separate content view of sensation of movement.As we move on, we will see that these three points become important.Theories of motor control assign roles to proprioceptive signals in the control of movement.In the next section, we will consider two types of theories: a motor command theory and the active inference theory (in a recent review, Floegel et al., 2023, call the two families of models for plant control models and perceptual control models, respectively).We will show that to fit the two data points, the motor command theory and the active inference theory are forced to choose different options at the choice point.

Functional role of signals in motor control
Afferent proprioceptive signals from sensors in muscles, joints, and skin are sufficient for a conscious sensation of movement.So are efferent motor signals from M1.We have described the sensations of movement by following the signals.In this section we focus on the role of the signals in motor control.

The role of proprioceptive signals in motor control
M1 and other cortical motor regions are central to planning of future movements, as well as online adjustment and control of ongoing movements.An example of planning of movements is when an agent withholds a specific movement until a certain stimulus appears or a specific period has elapsed but is otherwise ready to move towards the target (Cisek et al., 2003;Cisek and Kalaska, 2005).An example of online control of movements could be situations where an obstacle appears in a planned pathway of a fast reaching-movement (Gallivan et al., 2018).The obstacle can be detected by visual or somatosensory input or as perturbation detected by proprioceptive input.As a response, the motor signals to the arm are adjusted.There is a crucial temporal and contextual element to the role of proprioceptive signals in online control of reaching and grasping.As we will see, the role of afferent and efferent signals, as well as the role of cortical processing and conscious sensation depend on how the task context engages attention and task-goals.In the end, this cognitive dimension makes a modelling perspective necessary.
When it comes to understanding the roles played by proprioceptive signals in the online control of reaching and grasping, we can adopt either a bottom-up or a top-down perspective on the information processing.From the bottom-up perspective, the central question is the extent to which the control of a reaching or grasping movement requires cortical processing of fine-grained information projected from the bodily receptors to cortical regions.By contrast, from the top-down perspective, the central question is the extent to which cognitive mechanisms (attention, voluntary decisions, expectations, task-sets) exert an influence on low-level motor processing, such as spinal processes involved in reflex movements.
From a bottom-up perspective, it seems clear that some proprioceptive signals play important roles in online control of reaching and grasping without reaching the cortex.This can be observed in spinal processes like presynaptic inhibition, reciprocal inhibition, recurrent inhibition and 1b inhibition of spinal MNs.The networks in the spinal cord have been shown to be able to control some stereotypic movement patterns such as alternating flexion-extension movement (Grillner, 2006;Kiehn, 2016;Sherrington, 1913).Focusing on these spinal networks, it becomes a trivial matter that some proprioceptive signals play important nonconscious roles in motor control, assuming that cortical processing is a necessary condition for consciousness.
The bottom-up perspective provides a framework for understanding the cortical processing of proprioceptive signals.A well-studied system is the simple stretch reflex (Liddell and Sherrington, 1924).When a muscle is stretched fast, the muscle spindles will increase their firing rate.Part of the response to this stretch is a direct monosynaptic excitatory activation of spinal motor neurons, which makes the stretched muscle contract.When the muscle is contracted, the stretch of the muscle will be counteracted and hence decrease the muscle spindle firing.The stretch reflex is measured by electromyographic (EMG) activity from the muscles around the stretched joint.It can be observed that the response to a mechanical stretch is composed of three characteristic bursts of muscle activity.These are known as the short, medium, and long latency reflexes (SLR, MLR, and LLR).The initial SLR burst comes with a latency of about ~40-60 ms, the MLR with a latency of ~60-80 ms, and the LLR with a latency of ~80-120 ms (Diener et al., 1983;Evarts and Tanji, 1974;Zuur et al., 2009).Only the LLR is mediated via cortical reflex pathways.Consequently, proprioceptive signals processed via cortical pathways can exert an influence on on-going motor behaviour within 80-120 ms of processing.
Another line of experimental evidence comes from studies of somatosensory evoked potentials measured with EEG in humans.The first components of somatosensory processes are registered at approximately 140 ms after the stimulus onset (Auksztulewicz et al., 2012;Auksztulewicz and Blankenburg, 2013).A recent review has identified a so-called somatosensory awareness negativity signal with a time window from 125 to 180 ms after stimulus presentation (Dembski et al., 2021).This suggest that cortical processes can exert an influence on relatively simple aspects of motor control within 80-120 ms of processing time and that signals can become conscious after 125 ms.Given these results, there is a lower limit for processing times of around 125 ms before proprioceptive information becomes consciously accessible.This means that from a bottom-up perspective, for conscious sensations of movement to play a role in control of movements, a processing time of at least 125-180 ms is required.
Things become more complicated if we look at the processing from a top-down perspective.Several studies have shown that the size of even simple spinal reflexes induced with electrical stimulation of afferent nerves can be modulated by attention.Measured with EMG activity, it has been demonstrated that spinal reflexes can be directly influenced by conscious attention (Bonnet et al., 1997;Hale et al., 2003;Nakagawa et al., 2018).This effect is most likely mediated via a cortical mechanism that involves modulation of outgoing motor signals from M1.Evidence for this claim comes from Gandevia and Rothwell (1987) who showed that directing attention towards one of two muscles modulated evoked potentials measured from the attended muscle after cortical electrical stimulation.From a top-down perspective, it thus seems possible to consciously modulate even very fast reflex responses to external perturbations (for a discussion of what this means for our concept of a reflex, see Prochazka et al., 2000).
The reviewed studies clearly demonstrate the intricate entanglement of the afferent and efferent information flows involved in online motor control.Even proprioceptive feedback signals from movements might often reflect both afferent and efferent sources.From a bottom-up perspective, only some proprioceptive signals are cortically processed and can influence motor control as conscious sensations of movementbut only after a temporal delay of at least 125-180 ms.From a top-down perspective, consciousness can modulate afferent processing and the way it influences efferent processes at short timescales at a spinal level.The functional intricacies raised by the bottom-up and top-down perspectives make it difficult to give a simple and consistent physiological answer with respect to the cortical location, the neural networks, the temporal scales, and the function of conscious sensations of movement in the control of movement.Getting a clearer picture of the role of proprioceptive signals in the cortically based control of reaching and grasping requires a modelling perspective.

Motor command theories
By motor command theories of motor control, we include any model with the following general architecture (see Fig. 1A): Given a desired movement outcome specified in an "external" sensory format (say, grasping an object specified in terms of proprioceptive and visual sensations), a so-called inverse model takes as input the desired target state and an estimation of the current state of the effector and computes as output the optimal motor command signal (in an "internal" muscle format) that is used to innervate muscle contractions to produce the movement outcome via spinal alpha motor neurons (Jordan, 1996;Wolpert, 1997).The motor command theories also comprise a so-called forward model that takes the motor commands as input and computes as output a prediction of sensations.The current state of the effector is estimated by integrating sensory feedback from proprioceptive sensors with the motor commands (Sainburg and Schaefer, 2004;Wolpert and Kawato, 1998) and forward model predictions (Franklin and Wolpert, 2011;McNamee and Wolpert, 2019;Mehta and Schaal, 2002).The estimated state of the effector is compared with desired movement outcome of the effector.When the difference between the two is zero, the target has been achieved and the movement has been completed.This basic architecture can be (and has been) augmented and extended in various ways with extra comparators and informational loops to explain, for instance, fast error correction (Jeannerod, 1997;Pisella et al., 2000), self-recognition (Jeannerod, 2006), and sense of agency (Grünbaum and Christensen, 2020;Haggard, 2017).The motor command theories have many names.They are sometimes called optimal control theory (e.g., Todorov and Jordan, 2002), internal models (e.g., Jordan and Rumelhart, 1992), or plant control models (e.g., Floegel et al., 2023).They all have in common the existence of an efferent motor command signal, and for that reason we have decided to call these models "motor command theories".
Importantly in the present context, the current state of the effector is estimated by combining both afferent proprioceptive signals, efferent motor command signals, and predicted proprioceptive signals.Given the transmission time of the proprioceptive signals from the sensors in the hand and arm to the cortex (~80-120 ms), if the estimation of the current state relied only on the afferent signal, the estimation would always be significantly delayed compared to the actual current state of the effector (Franklin and Wolpert, 2011).This would be detrimental to online control of fast and fluent movement, for instance, when the system must correct the movement in response to perturbations (Desmurget and Grafton, 2000).The system can adjust for this delay by basing the current state estimation on a combination of afferent signals, efferent signals, and predictions from efferent signals.In normal voluntary movement, the estimated current state of the limb is consequently based on a combination of afferent proprioceptive signals and efferent motor signals (Fig. 1A).Experimental evidence, as reviewed by Wolpert et al. (1998) and more recently by Welniarz et al. (2021), indicate that the cerebellum plays an important forward modelling role using the motor signals to predict likely proprioceptive feedback.
Where should we locate the conscious sensations of movement in this kind of architecture?The answer seems to be relatively straightforward.The afferent and efferent signals showed to be sufficient for a sensation of movement are exactly the signals involved in estimating the current state of the limb.The estimated current state can give rise to or is associated with conscious sensations of movement (Fig. 1A).This motor control architecture allows us to explain the two data points, that is, the two sets of results highlighted in Section 2. First, in a situation with no actual movement but where proprioceptive sensors are activated by vibration, the estimation of the current state of the limb will be based only on the afferent feedback signal (Fig. 1B).This should also correspond to the situation where the limb is being passively displaced by some external force.Second, in a situation where subjects are paralysed and anaesthetized by a peripheral nerve block but motor signals are being sent to the spinal cord either because participants are trying to move (Fig. 1D) (Gandevia et al., 2006) or because of TMS stimulation (Fig. 1C) (Amassian et al., 1989;Christensen et al., 2010), the estimation of the current state is based only on the motor signals (and/or prediction derived from the motor signals).As the experiment by Gandevia and colleagues demonstrated, this situation is sufficient to produce a sensation of movement or at least a sensation of position.
With respect to the choice between a common or separate content view of sensation of movement for active vs passive movements, motor command theories are consistent with a separate content view.The distinctive contribution of efferent signals to the estimation of the current state in the case of active movement could explain why they feel different from passive movement.Importantly, this reliance on both corollary discharge signals and comparator signals differentiates this mechanism from the standard comparator model of sense of agency (Christensen and Grünbaum, 2018;Grünbaum and Christensen, 2020).Finally, it is worth mentioning that in a situation where a deafferented person performs a voluntary movement (Fig. 1E), we are still left with a small puzzle, because it is not entirely known what the forward model of this architecture will predict in terms of sensory feedback.The theoretical prediction depends on whether the forward model has been updated according to the absence of any proprioceptive feedback.Will the forward model predict complete sensory silence?Or will it infer movement based on the efference copy signal alone and predict some other form of sensory evidence?
To sum up, the motor command theory of motor control fits the two data points very well and is consistent with a separate content view.Assuming that the estimated current state of the limb is associated with conscious sensations of movement, we have the following three consequences.First, when afferent proprioceptive signals are the only input to the estimation, afferent proprioceptive signals can be sufficient for conscious sensations of movement (Fig. 1B).Second, when efferent motor signals are the only input to the estimation, efferent motor signals can be sufficient for conscious sensations of movements (Figs.1C and D).Third, the architecture is consistent with a separate content view.In fact, passive and active movements are plausibly associated with separate types of conscious sensations of movement depending on the absence or presence of motor signals.Associating conscious sensations of movement with the estimated current state does not imply that this estimation is always conscious or attentionally monitored.Notice also that it is consistent with the idea that fast error correction of movement happens outside consciousness if this correction is driven by a comparison of predicted feedback from the effector and the desired movement outcome of the effector (Desmurget et al., 1999;Pisella et al., 2000).This comparison might bypass the estimation of the current state.This fits with the assumption that the cerebellum plays an important role in forward model sensory prediction.The estimation of the current state of the limb and by extension the conscious sensations of movement most likely play important roles in motor learning and motor planning, even if they play no significant role in fast error correction.

Active inference theory
By active inference theories of motor control, we mean a theory with the following general form of architecture (see Fig. 2A): Given a predicted ("desired") perceptual state (say, grasping an object specified in terms of proprioceptive and visual sensations), the system computes the difference between the predicted perceptual state and the current perceptual state (Adams et al., 2013;Floegel et al., 2023;Friston, 2011;Shipp et al., 2013).The goal of the system is to minimize this difference or error.The system uses the error signal to predict the proprioceptive state that would best minimize the difference between predicted perceptual state and the current perceptual state.This proprioceptive prediction is transmitted to motor neurons in the spinal cord.Here the proprioceptive prediction is compared with the current state of signals from muscles and joints.If there is a difference, the error signal drives muscle activation until the error has been minimized.According to proponents of this type of motor control theory, a key advantage of the active inference theory over a motor command theory is that all informational transactions are kept in a perceptual format (Floegel et al., 2023).As a result, there is no need to postulate an inverse model that translates from a perceptual target state to motor signals ("external" to "internal" format) or a forward model that translates from motor signals to predicted sensory feedback ("internal" to "external" format).Proponents of the active inference theory sometimes appeal to simplicity considerations.According to the model there are no motor commands, just proprioceptive predictions (Friston, 2011).Another advantage of the model is that the general processing structure is the same at all processing levels.This also means that a prediction error at a higher level becomes the lower level's prediction while keeping the same "perceptual" signalling format throughout the entire process.
Where in this architecture (Fig. 2A) should we locate conscious sensations of movement?To answer this question, we need to consider the more general question of consciousness in the predictive processing architecture.The active inference theories of motor control are an implementation of a more general predictive processing theory of the brain (Friston et al., 2006;Hohwy, 2013;Seth, 2015).According to the predictive processing theory of information processing in the brain, the central goal of the brain (on all levels, from single neurons and networks to the whole brain) is to predict the sensory input by using a model of the underlying causes.The goal is to minimize the difference between the actual sensory input and the prediction using the error signal to update the model to make better predictions.Importantly, the goal of minimizing prediction error can be achieved in two ways.The system can either change its model used to estimate the predictions (so-called perceptual inference) or the system can alter the sensory input by moving to make the sensations match the prediction (so-called active inference).It is this last form of error minimization that provides an architecture for motor control (Floegel et al., 2023).
Despite not being a dedicated theory of consciousness, proponents of the predictive processing framework have offered explanations of consciousness (Seth and Bayne, 2022).According to one prominent version, perceptual consciousness corresponds to the brain's sensory predictions (Seth, 2021).Consciousness is not an afferent bottom-up signal from sensory receptors.Instead, consciousness is the top-down prediction signal based on the brain's best guess about the causes of the sensory signals.In short, perceptual and sensory consciousness is the brain's prediction of the sensory signal rather than the sensory signal itself.In terms of the predictive processing framework's preferred Bayesian formalism, sensory consciousness is the posterior belief (best guess) computed by multiplying the prior belief and the likelihood (the sensory evidence).If the guess is perfect, it means that the precision of one's the prior distribution is so high that multiplying it with the likelihood function produces a posterior almost identical to the prior.That is, new evidence makes no difference to one's belief.Sensory consciousness is the posterior, continuously updated and adjusted by the likelihood (sensory evidence) to minimize the error.Consequently, conscious proprioceptive sensations should correspond to a form of proprioceptive prediction (Adams et al., 2013;Floegel et al., 2023;Friston, 2011;Shipp et al., 2013).
Given the general predictive processing framework, we should locate conscious proprioceptive sensations in the proprioceptive predictions used to predict the proprioceptive signals from sensors in muscles, joints, and skin.In the active inference theory of motor control, the most likely candidate for the proprioceptive prediction is the signal flowing out from M1 (see Fig. 2A).Assuming this location for conscious sensation of movement in the architecture, how would the active inference theory of motor control explain the two data points?
First, in a situation with no actual movement but where proprioceptive sensors are activated by vibration (Fig. 2B), the system has the job of explaining the sensory input from the sensors.By using a model of the effector, the system predicts the proprioceptive sensations.The conscious proprioceptive sensations correspond to the brain's best guess (posterior).Consequently, the active inference theory of motor control would predict an outflowing predictive signal from M1 to the spinal cord, even in this situation with passive stimulation.The active inference model thus predicts activation of M1 when experiencing passive sensations of movement in the vibration experiments.This prediction is supported by results from Naito et al. (2002).Naito and colleagues found motor cortex activation in an fMRI study inducing illusion of movement by muscle vibration.Furthermore, the active inference model would predict that were the limb not blocked or restrained, the limb would move or drift to minimize the error between predicted sensation and actual sensation.This corresponds to the active inference explanation of drift in the rubber-hand paradigm (Maselli et al., 2022).To our knowledge, this prediction has not been experimentally tested in the case of muscle vibration.
Second, let us consider situations with M1 signals but no afferent proprioceptive feedback.In the situation with TMS induced illusion of movements (Fig. 2C) (Amassian et al., 1989;Christensen et al., 2010), there is an outgoing signal, which then should be interpreted as the proprioceptive prediction.Given the anaesthetized limb, there are no bottom-up sensations from the proprioceptive sensors to be explained.Total radio-silence from the sensors.In the situation where subjects are paralysed and anaesthetized, but motor signals are being voluntarily generated and sent to the spinal cord (Fig. 2D) (like the experimental situation of Gandevia et al., 2006), we have again what should be proprioceptive predictions from M1 but no afferent signals from the muscles.Again, there is radio-silence from the sensors.In both these situations, the proprioceptive prediction signal assumed to flow from M1 is not met by any proprioceptive confirmation at the level of the spinal cord.In these situations, it remains unclear how the active inference theory would explain the sensation of movement.How would the subsequent updating process go?It could either be a situation where the sensory evidence, i.e. the likelihood function, is an imprecise and wide (almost flat) distribution, in which case the proprioceptive prediction would remain unaltered by the evidence.This could explain the efference driven sensation of movement.Or it could be a situation in which the sensory evidence would be very easy to predict.The agent could simply refrain from moving.That is, the agent would predict no movement.This would be a motor version of the dark room problem for predictive processing (Sun and Firestone, 2020).If that were the case, the proprioceptive prediction should tend toward complete sensory silence and immobility.But this would leave us with no explanation of the sufficiency of efferent M1 signals to produce a sensation of movement.One option would then be to push the prediction up in the sensory hierarchy and claim that this type of proprioceptive prediction is generated by predictions flowing into M1 (maybe from premotor areas or prefrontal areas).At this stage, the active inference theory remains under-specified and open.The theory does not provide us with any clear explanations of the second data point.
In contrast to the motor command theory, the active inference theory entails an active version of the common content view of sensation of movement.The active inference model is struggling to explain the apparent difference between sensation of movement for active and passive movement (in terms of signal processing, passive movements are similar to the vibration induced illusion depicted in Fig. 2B).For both types of movement, the sensation of movement is explained as a proprioceptive prediction.That is, in both cases, we should predict an outflowing signal from M1 trying to close the gap between outflowing predicted sensation and inflowing sensory signal from proprioceptive sensors.
Finally, it is again worth mentioning the case of deafferented persons (Fig. 2E).Deafferented individuals should, according to the active inference theory, remain immobile.The brain's best prediction of the proprioceptive input should be no input at all.So, if motor signals from M1 are proprioceptive predictions, we should predict no movement.This was initially the case for case IW (Cole and Paillard, 1995).IW regained some capacity for voluntary movement.Given the active inference theory, IW's brain starts to produce proprioceptive predictions.In the absence of proprioceptive feedback this is puzzling.
To sum up, the active inference theory of motor control can provide several different explanations of the two data points depending on additional assumptions and specifications.Dependent on the situation, different parts of the network architecture would serve the purpose of providing the signals sufficient for sensation of movement.First, the active inference explanation of sensations of movement in the vibration studies is very different from the explanation by the motor command theory.For the active inference theory, afferent proprioceptive signals Grünbaum and M.S. Christensen should never be sufficient for a conscious sensation of movement.According to the general predictive processing framework, conscious sensation is associated with the downflowing proprioceptive predictions.The upwards flowing afferent signals from proprioceptive sensors should be met by downflowing proprioceptive predictions.Proprioceptive predictions should activate the motor system to close the gap between up flowing sensations and down flowing predictions.Second, the active inference explanation of sensation of movement generated by signals from M1 in anaesthetized and paralysed conditions remain under-specified at best, enigmatic at worst.If motor signals from M1 to muscles are proprioceptive predictions and the system's primary goal is to minimize prediction error, then in the case where there is no sensation from proprioceptive sensors, we might not expect persistent proprioceptive prediction of movement.Finally, assuming an active inference theory, passive and active movements are not plausibly associated with separate types of conscious sensations of movement (dependent on the absence or presence of motor signals).If sensation of movement is proprioceptive prediction and proprioceptive predictions drive muscle activation, then we should expect the passive movement to be like an active movement.The active and passive sensations should have a common type of content.
Assuming the data points (afferent signals are sufficient and efferent signals are sufficient) and the implication at the choice point (an active version of the common content view), we might have reason to prefer a motor command theory over an active inference theory.The motor command theory provides a more precise explanation of the sensation of movement and the experimental results.The active inference theory of motor control provides us with an under-specified explanation of the central aspects of the sensation of movement.In Popperian terms, this makes the theory less risky and less testable.The active inference theory has too much flexibility.If it does not add up, we could always place the proprioceptive predictions higher in the processing hierarchy (that is, higher up than the proprioceptive predictions flowing out from M1) and insist on some role for up-flowing signals in updating the posterior (the brain's best guess).But the active inference now seems to be confronted with a dilemma.Either it becomes too flexible and less testable, or it becomes a notational variant of the motor command theory.In either case, we have reason to prefer the motor command theory's explanation of the sensation of movement.As an example of this dilemma, consider a recent version of the active inference theory of motor control by Priorelli et al. (2023).The model is impressive in its ability to model even rather complex arm or whole-body movements.The model's ability comes at a price.Contrary to other active inference models, it works by assuming both a world-centred coordinate system and a body-centred coordinate system.The success of the model is due to the translation between the two coordinate systems, so intentions can be formulated in world-centred coordinates and translated into body-centred coordinates.Even if this model does not assume motor commands and an inverse model, it is arguably not simpler than any motor command theory.Rather than two different types of signals (motor and sensory), the model assumes instead two separate coordinate systems.
The flexibility of active inference theory might be a disadvantage with respect to understanding pathology and designing rehabilitation in patient groups characterized by abnormal experiences of voluntary movement, such as patients with schizophrenia (Frith, 1987;Daprati et al., 1997), functional neurological disorders (Hallett et al., 2022), Parkinson's disease (Amanzio et al., 2010), cerebral palsy (in particular with dystonia), and neuropsychological disturbances following stroke including anosognosia for hemiplegia (Piedimonte et al., 2016).These disorders are partly characterized by altered sensations of movement.The two different motor control architectures could have very different implications for how to understand the causes of the impairment and how to address the impairments.First, depending on neural localization of parts of the functional architecture, we can use lesions to make predictions about altered experiences and we can use altered experiences to make predictions about possible lesions.If a motor control theory cannot make sufficiently precise predictions about the consequences of a lesion to a specific brain area, it is less useful as a tool for understanding the pathology.Here the motor command theory seems to be better at making precise predictions about symptoms.Second, one crucial factor in neurorehabilitation is that patients are actively engaged in the rehabilitation process (Nielsen et al., 2015).The fact that the motor control theory distinguishes between active and passive movement might make it easier for the model to explain the mechanisms underlying the known benefit of active participation in rehabilitation.

Are proprioceptive signals introspectively accessible?
Up until now, we have focused on the signals involved in conscious sensations of movement.We have seen that these signals play welldefined roles in motor control, assuming a specific architecture for motor control.We have assumed that we can study the functional role of conscious sensations by studying the role of the signals.This assumption is controversial.The vibro-tactile studies show that under some conditions the signals are introspectable.This does not establish that the signals are introspectively accessible also during reaching and grasping.The fact that the signals play an important role in the online control of movement does not suffice to establish the role is played by conscious sensations of movement.Maybe the strategy of simply following the signals do not lead us to conscious sensation of movement.To address this question, we need to review studies directly studying participants' ability to monitor their proprioceptive signals during movement.

No access to proprioceptive signals during reaching and grasping
The general picture suggested by numerous studies of perceptual processing suggest that sensory signals from a task relevant modality can play important roles in action control even when the signals are not consciously processed.Related studies have been conducted in the domain of proprioceptive processing.Several studies imply that cortically processed proprioceptive information plays a role in online adjustment without being consciously accessible.
The dissociation between sensory signals processed non-consciously for motor control and sensory signals processed consciously for cognition has been explored in the visual domain.Perhaps the best-known studies concern cases of blindsight and visual form agnosia (Goodale and Milner, 1992;Sanders et al., 1974; for a TMS induced equivalent of blindsight, see Christensen et al., 2008).Studies of blindsight typically involve patients with lesions in primary visual cortex.These patients, when tested with visual stimuli presented in their blind visual field, can perform some motor actions correctly in relation to the presented visual stimuli, while being unable to report what the visual features of these stimuli are.The standard explanation is in terms of a dissociation between visual information feeding into the control of a movement and visual information processed for conscious perception.Even if the visual information is not consciously perceived, the visual signals are still able to play a role in the motor control processes enabling the patient to reach for and grasp a target.
In a related type of experiments, neurologically intact participants perform reaching movements guided by visual signals in the absence of visual signals from the moving limb (Pélisson et al., 1986;Sergio, Scott, 1998).In these experiments, participants were unable to report a change in the visual location of a double step target shift, but movements of the arm were corrected and performed correctly to the new target position.An interesting version of this online target tracking task was an experiment where participants performed a pointing movement towards a stationary target, but if the target changed position, they had to refrain from correcting their movement to the new target location (Pisella et al., 2000).Despite this relatively simple instruction, some participants found it difficult to refrain from making the online correction.This suggests that even if the participants are aware of the target jump, the conscious aspect might not play any significant role in the way the system uses the visual information to guide the hand to the new target location.Though controversial (Grünbaum, 2017;Phillips, 2021;Shepherd and Mylopoulos, 2021;Wu;2020), these studies are often taken as evidence for the dissociation between the role visual signals play in online control of reaching and grasping and the role they play in conscious visual perception (Clark, 2001;Jacob and Jeannerod, 2003;Milner and Goodale, 1995).Object-directed motor control appears to work just fine in the absence of conscious visual perception of the objects but of course not in the absence of visual information.
To directly study the function of conscious sensation of movement in motor control, we need another type of experiment that focuses on sensation of movement rather than visual experience.Fourneret and Jeannerod (1998) famously tried to dissociate the role of proprioceptive signals in motor control from their role in conscious sensation of movement.They asked participants to perform a simple straight-line drawing movement with veridical or manipulated visual feedback.The actual movement of the arm was hidden to the participants, and they could only see the visual indication of the movement.The main finding of the manipulated feedback trials was that participants adjusted their movement in the opposite direction of the perturbation introduced by the manipulation without being able to report verbally that they had performed the adjustment.This led the authors to suggest that humans have only limited conscious access to the proprioceptive feedback from their motor performance.Signals concerning the actual location of the hand, its movement direction, velocity, and force, obviously played important roles in allowing the participants to adjust their movements to correctly satisfy the task instruction of drawing a straight line.That is, the proprioceptive information played its normal roles in the online control of the reaching movement, but it did this while participants apparently remained unaware of the actual limb location and movement direction.
The basic structure of this type of paradigm is to give participants a reaching or grasping target, while manipulating sensory signals or perturbing the movement and measuring compensatory strategies.In some experiments, researchers have applied a force field to a robot arm or handheld manipulandum.To satisfy the task instructions and reach for the target, subjects typically adjust their muscle activity without noticing the adjustments (Franklin and Wolpert, 2008).More recent studies have used virtual reality to increase the sense of embodiment and ownership (see, for instance, Padrao et al., 2016).In one study (Maselli et al., 2023), experimenters used a realistic virtual model of the participants' hands.The task was to reach for a target in the virtual space where reaching the target required participants to adjust the movement either away or towards the body midline.After reaching the target, the participants performed a two-alternatives-forced-choice task where they indicated whether the reaching movement had felt normal or weird.Results indicated that when participants were most efficient at correcting for perturbations (towards the body midline), they tended to not notice their motor adjustments (they classified the performance as normal) (Maselli et al., 2023).Using a different virtual reality task, Bourdin and colleagues had participants make large-scale adjustments of arm movements that apparently went unnoticed (Bourdin et al., 2019).Participants had their arm in an elastic band and were instructed to make movements of the lower arm of 90 • .In the virtual world, the movements were transformed such that participants had to either adjust by 0 • , -15 • , or + 15 • to reach the target of 90 • arm flexion in the virtual world.Throughout the experiments, the participants were instructed to visually attend to their virtual hand.According to Bourdin et al., the results provided "direct evidence that, at least under specific experimental conditions, participants can significantly modulate their movement amplitude and muscle activity to accomplish a task, being completely unaware of such accommodation" (p.4).
Some patient studies also support a dissociation between conscious sensations of movement and proprioceptive signals in control of reaching and grasping.For instance, Wimperis and Wing (2007) tested a patient who was unable to report tactile perturbations of a grasp movement while still being capable of performing the grasp correctly.
Reflecting on this type of data, Dijkerman and De Haan (2007) have proposed a processing architecture for the somatosensory domain similar to the dissociation between vision for perception and vision for action (Goodale and Milner, 1992;Milner and Goodale, 1995).They argue that the anterior (APC) and posterior parietal (PPC) cortices display a similar split as the dorsal and ventral streams, with APC responsible for proprioceptive perception and PPC for proprioceptively guided action.
The overall impression from this line of experimental studies seems to be that conscious sensation of movement plays no vital role in the online control of movements.Collectively, these studies have been taken to show that "awareness of a movement is not influenced by the sensory output generated by the movement" (Bourdin et al., 2019, p. 4).Consequently, results from studies using reaching and grasping tasks indicate that the signals involved in the estimation of the current state of the effector during movement are not associated with a conscious sensation of movement.Reflecting on these results, we might reach the conclusion that when using the strategy of following the signals, we do not arrive at conscious sensation of movement.

Metacognitive access to sensations of movement
The data about the introspective accessibility of proprioceptive signals are contradictory.In our discussion of afferent proprioceptive signals, we reviewed studies using muscle vibration that demonstrate that when we artificially activate the proprioceptive receptors in the muscles in ways similar to their activation by natural movement, subjects have sensations of movements.The sensations have content that is sufficiently fine-grained to allow the subject to track the movement direction, velocity, and force.But in Section 4.1.,we reviewed a second set of studies using reaching and grasping tasks that seems to demonstrate that we have no introspective access to the proprioceptive signals.Even relatively large-scale adjustments of movements seem to go unnoticed if movements visually appear to reach their target.These would not be the results we would expect if all movements were accompanied by relatively fine-grained conscious sensations of movement.We will now survey a third type of studiesstudies of metacognitive access to proprioceptive signalsthat appear to show that participants in experiments employing reaching tasks do in fact have introspective access to the proprioceptive signals.
Recently, several studies have investigated subjects' metacognitive access to proprioceptive signals by using a design developed to study the role of metacognitive assessment in visual discrimination tasks.One measure of interest when studying people's ability to visually discriminate between pairs of stimuli is their metacognitive assessment of their performance in the discrimination task.For instance, participants could have the type 1 task of discriminating between two possible orientations of a Gabor grating (horizontal vs vertical) and on each trial have the additional type 2 task of providing a confidence rating of their type 1 task performance.Whereas the type 1 task is to discriminate between external stimuli provided by the experimenter, the type 2 task is to discriminate between one's own correct and incorrect decisions (Clarke et al., 1959).The ability to metacognitively assess one's own decisions has long been acknowledged to be important for learning from experience and improving one's decision-making (Galvin et al., 2003).More recently, type 2 confidence rating has been used as a measure of introspective access and consciousness (Dienes and Seth, 2010;Fleming and Lau, 2014).
Signal detection theory has for a long time been used to provide a bias free measure of the participants discriminatory sensitivity in type 1 visual decision task (Green and Swets, 1966).More recently, researchers have begun developing bias free ways of analysing metacognitive sensitivity when discriminating between correct and incorrect decisions (Kunimoto et al., 2001;Persaud et al., 2007).Metacognitive sensitivity is a measure of subjects' reliability in discriminating between their own T. Grünbaum and M.S. Christensen correct and incorrect decisions (that is, how good they are in detecting their own correct and incorrect type 1 judgements).Roughly, it is a measure of the accuracy-confidence correlation.
Two problematic confounds are important to assess when describing metacognitive sensitivity (Fleming and Lau, 2014;Maniscalco and Lau, 2012).First, sensitivity can be affected by the type 1 performance.The difference between a participants' metacognitive sensitivity in two different tasks might be explained by difference in task difficulty.This confound is often avoided by using staircasing procedures to ensure an equal level of task difficulty across conditions and participants.Second, metacognitive sensitivity can be affected by a response bias in the type 2 decision.One person might give higher confidence rating than another person simply because they are more ready to apply the confidence categories.To avoid this second confound, Maniscalco and Lau (2012) developed a model-based approach to estimate what they termed meta-d′.Meta-d′ is the d′ for the type 1 decision an ideal agent should have, given the observed type 2 ratings.That is, meta-d′ is the d′ an agent that is optimally using all available information for the type 2 judgement would have, given the observed type 2 responses.By relating meta-d′ to the actual d′ for type 1 task, it is possible to estimate the participants' general over-or under-confidence.The meta-d′/d′ ratio is often taken as a bias-free measure of metacognitive efficiency (Maniscalco and Lau, 2016).If the ratio is close to 1, it means that the information available for making the type 1 decision is most likely also available for making the type 2 judgement.
Recently, researchers have been using versions of this method for assessing metacognitive sensitivity in simple motor tasks.Sinanaj et al. (2015) used a simplified version of the reaching task by Fourneret and Jeannerod (1998).The visuomotor task was to draw a straight line to a target by pushing forward a joystick handle, without direct vision of their hand.Visual feedback of the trajectory was provided in real-time on a computer monitor.During the reaching movement a deviation was introduced to the line.To reach the target, participants therefore had to adjust their reaching movement.After each trial, participants had to decide (yes/no) whether they detected any deviation of the movement and then rate their level of confidence in the accuracy of their decision using a five-point Likert-like scale.Deviations occurred in 79 % of all trials.To ensure a similar level of performance in the type 1 detection task across participants, a staircasing procedure was used by adjusting the degrees of deviation to ensure an accuracy rate between 71 % and 74 % for detection of deviation in all deviation trials.In a similar study from the same lab (Pereira et al., 2023), the researchers assessed how well information from the type 1 detection performance informed the metacognitive type 2 judgement.The authors calculated the meta-d′/d′ ratios for trials with "no" responses (0.96) and trials with "yes" response (0.99).Not only does this indicate that evidence available for making the detection decision was available also for making the confidence assessment.The results also indicate that even on those trials where participants failed to report deviations, their confidence judgements still discriminated reliably between deviated and non-deviated trials.
The design of the studies by Sinanaj et al. (2015) and Pereira et al. (2023) makes it impossible to determine the role of proprioceptive signals in the metacognitive processing and judgement.The metacognitive sensitivity of the type 2 assessment could be drawing on several types of informational sources.Cues related to goal-satisfaction, perceptual accuracy, and proprioception, as well predictive signals related to motor commands and intentionsnot to mention general knowledge and perceptual expectationscould all lie behind a high accuracy-confidence correlation (Locke et al., 2020).To tease apart some of informational sources for the confidence ratings, recent studies have compared metacognitive sensitivity across comparable visual, visuomotor, motor, passive, active versions of the same basic task.
In a study by Charles and colleagues (Charles et al., 2020), participants had to perform type 1 and type 2 decisions comparable across an active, a passive, and a visual condition.In the active condition, the participant's finger was attached to a robotic arm capable of recording finger position.Participants made back and forth movements between an upper and a lower bound while their finger position was displayed as a green dot on the screen.At a crucial phase of movement, the finger position became invisible during which a probe (blue dot) appeared while a brief tone was played.The type 1 response was to decide whether the probe had appeared ahead or behind of their finger position.The type 2 decision was to report their confidence in their response using a 4-point scale.Passive condition was similar, except the robotic arm passively moved the finger of the participant up and down.The study found relatively high metacognitive sensitivity across the conditions and no significant difference between meta-d′/d′ ratios for the three conditions.To the extent that access to proprioceptive signals explains the type 2 performance in the passive conditions, it would seem fair to assume that they also play a role in the active condition.
Relatedly, Arbuzova et al. (2021) tested metacognitive sensitivity in a simple motor task across motor, visuomotor, and visual conditions.In their basic motor task, participants threw a virtual ball that they released from a moveable lever.In their first experiment, the lever and the ball's trajectory were shown on the screen, except in the motor condition where only the lever was shown.The participants' type 1 decision was to choose which one of two trajectories displayed on the screen immediately afterwards corresponded to their own ball throw.Participants then had to rate their confidence in their decision.In the visuomotor and visual conditions, participants were shown the trajectory of the ball on the screen.In the visual condition, the participant made no arm movement.In the second experiment, the basic motor task was identical to the first experiment.In the motor condition, the trajectory of the ball was shown but not the lever.Now the type 1 decision was which of two angles of the lever displayed on the screen corresponded to their arm position at the time of the ball release.Finally, they rated confidence in their decision.The aim with this set-up was to ensure that participants in the motor condition used proprioceptive information to make their type 1 decisions.To decide the angle of the arm at the time of releasing the ball, they had to rely partly on proprioceptive information from the arm.Across all conditions of both experiments, the mean meta d′/d′ ratios were well above zero indicating that participants used information from their type 1 judgement to make their type 2 assessment (they were not guessing).The analyses also showed that there were no significant differences between the mean ratios in the different conditions.We can therefore assume that the information available in the motor condition is also available in the other conditions.
Reflecting on studies of metacognitive sensitivity in motor tasks and motor cognition, Arbuzova and colleagues (2021) conclude that the experimental results contradict "the view that minute parameters of own movements are not available to awareness and humans predominantly monitor the outcome of the movement (whether its goal has been achieved) or its effect in the world" (Arbuzova et al., 2021(Arbuzova et al., , p. 2224)).The results thus indicate that people can "accurately monitor their voluntary movements in the absence of corresponding visual cues" (Arbuzova et al., 2021(Arbuzova et al., , p. 2225)).To be sure, the current experimental set-ups do not allow us to tease apart which motor related signals people are using in their type 1 and type 2 decisions in the motor conditions.Participants may primarily rely on efferent signals, afferent signals from proprioceptive receptors, or (perhaps more likely) a combination of both.Importantly, the results indicate that people have introspective access to the type of information involved in sensations of movement (location of the limb as well as its movement direction, velocity, and force).We cannot explain the results simply by the brain's monitoring of visual signals.The motor conditions exclude the visual component without significantly altering the metacognitive sensitivity.

The paradox of sensation of movement
Let us assume that we have conscious sensations of movement only if the proprioceptive signals are introspectively accessible.The implication is that the results from the reviewed studies seem to support conflicting predictions.Some results from studies where participants reach for visual targets support the claim that during reaching and grasping, subjects have no significant introspective access to proprioceptive information from the arm movement.Within certain limits, arm movements can be manipulated without participants noticing, so long as the visual target is reached (Slachevsky et al., 2001).By contrast, the results from vibration experiments and the metacognitive paradigms support the claim that during reaching and grasping subjects have introspective access to fine-grained proprioceptive information from the arm movement.Both claims cannot be true, yet both seems to be supported by sound empirical evidence.
We cannot brush away this paradox simply by saying that data from visual reaching tasks and vibration experiments cannot be compared directly because they embody very different attentional demands.According to this simple "attentional" solution to the paradox, the different paradigms provide us with results that point in different directions simply because the experimental tasks put very different demands on participants' attentional and cognitive resources.In the vibration experiments, participants are asked to attend to their proprioceptive sensations.By contrast, in most reaching tasks, participants are instructed to attend to a visual target.If we assume that consciousness and reportability requires attention (Dehaene et al., 2006;Dehaene and Changeux, 2011;Dehaene, 2014), we should conclude that proprioceptive information during reaching and grasping paradigm is unattended and therefore not conscious.
This explanation of the conflicting results is problematic for several reasons.First, data from vibration experiments are often supposed to tell us something about the type of proprioceptive processing that occurs generally in normal reaching and grasping.All normal reaching and grasping activate the proprioceptive receptors.When information from proprioceptive receptors reaches the cortex, it gives rise to the sensation of movement with a representational content sufficiently fine-grained to allow participants to shadow the limb's position, movement direction, velocity, and force (Naito, 2004).This information should also be accessible when people make reaching actions.Second, the metacognitive experiments use versions of the task of reaching for a visual target.In some cases (Sinanaj et al., 2015;Pereira et al., 2023), the behavioural tasks are designed to be directly comparable to the task used by Fourneret and Jeannerod (1998).Results indicate that people are metacognitively sensitive to sensations of movements.This gives us reason to think that proprioceptive information about movement is introspectively accessible also during reaching and grasping.
The motor command theory of motor control offers us a way to explain the results of the vibration experiments and the reaching tasks in a consistent manner.We assume that conscious sensation of movement is associated with the estimation of the current state of the limb.In voluntary movement in the intact agent, this estimation is based on an integration of afferent signals from proprioceptive sensors, efferent motor signals to muscles, and proprioceptive sensory predictions from the forward model.Normally, the estimated proprioceptive state of the limb is associated with a conscious sensation of movement, even if the conscious sensations of movement can be recessive, fleeting, and hard to report, for instance, because conscious attention is tied up elsewhere.In addition, this architecture allows us to accept that sometimes the estimated proprioceptive state of the limb is non-conscious.We can imagine noisy or attentionally demanding situations in which the proprioceptive signal is masked or weakened.Importantly, we can also computationally explain situations where the predicted proprioceptive sensation from the forward model plays important roles in fast error correction and online control of perturbed movements (Desmurget and Grafton, 2000).In these cases, the prediction signal bypasses the state estimation and is used directly in online control.If conscious sensation of movement is associated with the estimation of the current state of the limb, we can allow that the prediction is used non-consciously in online control of action.Notice that this type of explanation is not open to the active inference theory of motor control, according to which the prediction of proprioceptive sensation is the conscious sensation (the brain's best guess of the causes of the sensation).

Concluding remarks
In this paper, we have focused on the type of signals that are sufficient for a conscious sensation of movement.To describe the role of conscious sensation of movement in the control of movement, we adopted the strategy of following the signals established to be sufficient for conscious proprioceptive sensations.By reviewing studies using muscle vibration, we highlighted the fact that afferent signals from activated muscle spindles are sufficient to generate a sensation of movement with a specific orientation, direction, velocity, and force.By discussion of studies using a peripheral nerve block, we highlighted the fact that also efferent signals from M1 are sufficient to generate a sensation of movement, or at least a sensation of position.Together, these results are consistent with the claim that active sensations of movement and passive sensation of movement are phenomenologically and physiologically very different.The distinctive contribution of efferent signals in active movement supports this idea.By locating the signals known to be sufficient for a sensation of movement in the architecture of a motor command theory (such as optimal control theory), we showed that the conscious sensation of movement most likely corresponds to the estimation of the current state of the limb (estimated based on afferent, efferent, and predictive signals).By contrast, if we locate the sensations in the active inference architecture, the sensations should correspond to proprioceptive predictions.Proprioceptive predictions most likely correspond to efferent signals from M1.The two theories and their respective architectures have strikingly different implications for the explanation of the functional role of sensation of movement.
Adopting the strategy of following the signals known to be sufficient for a conscious sensation of movement has important implications for our understanding of the role of sensation in motor control.We hope we have demonstrated that the study of conscious sensation of movement holds an important key to opening and addressing central problems in research on motor control and consciousness.