During motor learning, humans rely mainly on visual information to produce accurate and stable behaviors (for a review, see Wolpert et al., 2011). According to the specificity of practice hypothesis (Tremblay and Proteau, 1998), although visual feedback initially helps to improve performance, it can lead to forms of dependency. As a result, when this feedback is removed, performance in delayed retention collapses. Studies have shown that in addition to movement coding based on visual cues (i.e., visuospatial coding), participants engage either in motor coding (Hikosaka et al., 1999, 2002) based on proprioceptive cues, or in a combination of the two (Kovacs et al., 2009). Moreover, in the case of unreliable visual feedback during the learning of a discrete task, they rely on other sensory sources (Körding and Wolpert, 2004; Berniker and Kording, 2008; Hewitson et al., 2018), as predicted by the Bayesian approach to multisensory integration (Ernst and Banks, 2002). Most of the work on the motor learning process has involved the use of discrete tasks, even though most of our daily actions are sequential (e.g., including several sub-tasks such as walking or brushing teeth; Hikosaka et al., 1999). In the present study, we therefore assessed the impact of on-line visual feedback reliability on the learning of a motor sequence.

Sensorimotor learning requires individuals to identify the source of information most likely to ensure maximum success on the task (specificity of practice hypothesis; Tremblay and Proteau, 1998). They do this early in practice, and rely on that information thereafter, to the detriment of other sensory sources. If a motor task requires high visuospatial accuracy, for example, visual feedback is rapidly established as the dominant source (Tremblay and Proteau, 1998). However, if this information subsequently becomes unavailable, the planning and control of movement will be disrupted, resulting in major aiming errors. Most of the studies testing the specificity of practice hypothesis have focused on the manipulation of task constraints (e.g., practice conditions or target sizes), mainly for discrete tasks (e.g., Robin et al., 2004; Proteau, 2005). However, Blandin et al. (2008) focused on the learning of a continuous arm flexion-extension sequence, showing that whereas the on-line visual feedback of the effector’s position was initially beneficial (i.e., during acquisition), it became detrimental during retention and transfer, if participants no longer had access to it. This was an important result, as it validated the specificity of practice hypothesis (Tremblay and Proteau, 1998) for sequential tasks and confirmed the evolving role of feedback across the different stages of learning (for a review, see Wolpert et al., 2011).

Moreover, the representation of motor skills relies on distinct and independent coordinates or coding systems (Coello et al., 1996; Criscimagna-Hemminger et al., 2003; Lange et al., 2004). Each coding system contributes to movement production and can produce specific learning and transfer capacities. According to Hikosaka et al. (1999, 2002), the learning of a motor sequence is encoded in a neural-network schema in two ways (i.e., visuospatial and motor coding). The spatial locations of the effector and the targets to be reached are coded in visuospatial coordinates. This coding, the dominant representation early in practice, is explicit (i.e., humans consciously use relevant visual information such as the positions of their limbs and the stimuli), and effector-independent (i.e., not specific to an effector). Additional practice leads to motor coding: motor patterns (i.e., activation of agonist/antagonist muscle) generate muscle activations and joint configurations specific to the effector producing the movement. Several studies have confirmed this initial dominance of visuospatial coding (e.g., Park and Shea, 2002, 2005; Kovacs et al., 2009; Boutin et al., 2010, 2012), and the relative underuse of motor coding (e.g., Kovacs et al., 2009; Boutin et al., 2010, 2012) through intermanual transfer tasks. During these classical transfer tasks, participants perform the sequence with their non-dominant hand and either the same targets’ sequence to point (which maintains the visuospatial coding; visuospatial transfer test, TVS) – or the mirror of this targets’ sequence (which maintains the motor coding; motor transfer test, TM). A higher performance in TVS would imply a higher visuospatial coding while a higher performance in TM would imply a higher motor coding (e.g., Park and Shea, 2002, 2005; Kovacs et al., 2009; Boutin et al., 2010, 2012).

The reliability of visual information has recently been identified as a potential factor influencing coding processes and, more particularly, promoting the implementation of motor coding (Robin et al., 2004; Hewitson et al., 2018): when visual information is unreliable, participants rely on proprioceptive information and undertake more motor coding. This hypothesis is based on the work of Körding and Wolpert (2004) who showed that participants exposed to visuomotor bias adapted their pointing movements relative to the level of uncertainty of the sensory feedback. In this study the feedback reliability was manipulated by providing the cursor position through the presentation of multiples dots distributed as a two-dimensional gaussian depicting a cloud of dots. Participants learned the sensory feedback likelihood over practice trials (Sato and Kording, 2014). The weight assigned to each sensory modality therefore depends on its relative reliability for the participant (Ernst and Banks, 2002).

In the present study, we used a continuous dynamic arm movement task to examine motor sequence learning according to the reliability of the on-line visual feedback. We specifically assessed whether coding of the task depends on this reliability, which was not already shown to our knowledge. We choose to manipulate the feedback reliability with a paradigm similar to Körding and Wolpert (2004) and Hewitson et al. (2018) which depicted a cloud of dots. Therefore here three conditions of reliability were tested: reliable (σ0), Moderately unreliable (σM), and Largely unreliable (σL). We predicted that the more unreliable the visual cursor, the harder the sequence would be to learn, as shown for discrete tasks. Because many learning experiments reported that unfavorable practice conditions usually favor learning (Sigrist et al., 2013 for a review), we first had to assess the effect of feedback reliability on learning. Moreover, we predicted a more efficient motor coding would take place in the two unreliable conditions than in the reliable condition. Indeed, as participants who could not trust the visual information would rely more on proprioceptive information and favor motor coding.

Figure 3 depicts MT per block as a function of the three phases of the experiment (i.e., acquisition, retention & transfer) for the three cursor groups (i.e., σ0, σM, σL). The following analyzes characterized (i) the evolution of performances in the acquisition and retention phases; (ii) the sequence-specific learning and (iii) the long-term learning and coding; first for the reliable condition and then for all conditions.

Our main objective was to assess the effect of visual cursor reliability on the learning and movement coding of a continuous motor task. First, our analyses revealed an improvement in performance for the reliable group (σ0) that could not result solely from visuohaptic calibration (i.e., set up and task familiarization): MT was shorter for the repeated sequence than for the new sequence in the posttest (i.e., R₂₀ vs. N₂ comparison). This typical learning process for the reliable condition was confirmed by a gradual decrease in MT during acquisition (i.e., power-law fitting). By contrast, our results showed that unreliable visual cursor disrupts the long-term learning of a sequence, for although we observed retention of performance for σL after 24 h (i.e., R₀ vs. RET_C), MT remained longer than for σ0. Finally, analyses of transfer yielded the classic finding of better performance on TVS than on TM for σ0, but no such difference was observed when the cursor was moderately (σM) or very (σL) unreliable. Interestingly, these results indicated for the first time that the type of coding of the sequence depends on the reliability of the cursor.

In line with previous results (e.g., Park and Shea, 2005; Blandin et al., 2008; Kovacs et al., 2009; Boutin et al., 2010, 2013a, 2014), we expected to observe a gradual decrease in MT across the acquisition phase. Here, we observed an overall decrease in MT across the blocks, whatever the reliability of the cursor (see power-law fittings). MTs were longer for σM and longer still for σL, compared with the reliable group (σ0), owing to the noise generated by the search for the reference dot, which particularly disturbed participants’ movements at the beginning of practice (see representative trajectories Figure 1C). Conversely, their MT decreased with practice. To assess the learning effect, we used the classic method of including a control block (e.g., Park and Shea, 2005; Boutin et al., 2014) to differentiate between the general improvement (i.e., generalization of practice) and sequence-specific learning (for a review, see Abrahamse et al., 2010). In the reliable group (σ0), in accordance with Boutin et al. (2010, 2013b, 2014), we observed a higher MT for N₁ relative to N₂ and a higher MT for N₂ relative to R₂₀, indicative of sequence-specific learning, but no such decrease was observed in the unreliable cursor’s groups. To our knowledge, this is the first time that unreliable visual cursor has been shown to lead to a decrease in learning performance on a sequential task. As visual information was available throughout the experiment, performance remained optimum for σ0, which may even have developed visual dependency (Blandin et al., 2008). By contrast, when the cursor was unreliable during acquisition (i.e., for σL and σM), participants performed more poorly than controls (σ0) on retention tests, even when there was a reliable cursor (i.e., RET_C). We hypothesized that participants in σL and σM learned to rely more on other sensory sources (e.g., proprioceptive cues), as these were more reliable during acquisition (Körding and Wolpert, 2004; Berniker and Kording, 2008; Hewitson et al., 2018), and they continued to use these sources even when they could benefit from reliable on-line visual feedback. One may argue that this reweighting toward other sensory sources may also be linked to the change of the visuomotor mapping during σL and σM conditions. Indeed, the cloud of dots may not only alter the reliability of the visual cursor, but also modify the visuomotor mapping between the hand and visual cursor positions. Participants would have been required to learn the stochastic visuomotor perturbation as the reference dot which validated the target had changed every 12 targets. Adaptation to visuomotor perturbations was shown to lead to slower and less complete learning when the standard deviation of the visuomotor rotations distribution was higher (Fernandes et al., 2012).

To observe the coding process, it is common to perform delayed visuospatial and motor transfer tests (e.g., Kovacs et al., 2009). For TVS, participants had to produce the same sequence as during acquisition or retention (i.e., repeated sequence), but with the contralateral hand: the visuospatial pattern was thus maintained, but the pattern of muscle activation and the joint angles were opposed. For TM, participants had to produce a mirror sequence using the contralateral hand: the pattern of muscle flexion and extension was thus the same as during the repeated blocks, but the visuospatial pattern was mirrored. For the group with reliable cursor (i.e., σ0), we confirmed the results of previous studies reporting longer MTs for TM than for TVS (e.g., Kovacs et al., 2009; Boutin et al., 2010, 2012): at this stage of learning, visuospatial coding was therefore more developed than motor coding. Furthermore, we first showed a link between cursor reliability and movement coding. For groups with unreliable cursor (i.e., σM and σL), we observed similar MTs for both TVS and TM. Thus, when participants initiated visuospatial coding, they also initiated motor coding, unlike those in the reliable group. This link was supported by the computation of the deterioration rates for TM ([MT of TM – MT of RET_C]/MT of RET_C) and for TVS ([MT of TVS – MT of RET_C]/MT of RET_C). Whereas σL and σM had similar deterioration rates for both TM (0.14 for σL and 0.13 for σM) and TVS (0.16 for both groups), σ0 had a greater deterioration rate for TM (0.31) than for TVS (0.19).

The nature of movement coding depends on task’s context and individual’s strategy (Blouin et al., 2014), as well as practice’s conditions (Hikosaka et al., 1999, 2002) suggesting a binary process. This means that if visuospatial coding was the dominant type at the start of the acquisition phase, motor coding could only replace it as a result of practice. However, Kovacs et al. (2009) found that TVS remained better than TM regardless of the amount of practice (i.e., 1, 4, or 12 days). To the best of our knowledge, our results indicated, for the first time, that the reliability of sensory information during training influenced the type of coding that was implemented: the more available and reliable the visual information during acquisition, the greater the implementation of visuospatial coding. By contrast, when this cursor was degraded, participants relied more on other available sources of sensory information, and engaged in both visuospatial and motor coding. This mixed coding showed that, depending on the context, one or other type of coding could be activated (Kovacs et al., 2009). One cannot exclude that this mixed coding hypothesis could be rather a unique coding/reference frame which would include more or less weight to the visual and proprioceptive cues reliable in this pointing task. The resulting multisensory coding would obey the “reliability rule” (Colonius and Diederich, 2020): the weights in averaging the unisensory estimates would be proportional to their relative reliabilities. Several studies showed that the perceived positions of the hand (conveyed by proprioception) and the object to reach (conveyed by vision) are biased toward each other (e.g., Rand and Heuer, 2013; Debats et al., 2017). This perceptual attraction has been shown to be modulated by the reliability of the hand position and the visual object position reliability (Debats et al., 2017). The brain would combine sensory sources in an optimal fashion providing an amodal unique representation/coding (Kirsch et al., 2016).

Overall, this study assessed the impact of the reliability of on-line visual feedback on sequence learning and how the motor sequence was coded. Although unreliable visual cursor had a negative impact on retention, it also stimulated mixed coding (visuospatial plus motor coding), promoting better sequence learning. As mentioned by Wulf et al. (2010), the most beneficial conditions for learning in the short term (i.e., acquisition phase) are not necessarily the most beneficial in the long term for effective learning as retained or generalizable skill or knowledge. Therefore, here the beneficial impact of visual unreliability on motor coding constitutes an interesting output, notably in the domain of education, performance or reeducation. Interesting perspectives also appear regarding the manipulation of the visual feedback reliability over several consecutive days of practice as Kovacs et al. (2009) showed that coding evolves over short to long term practice (1–14 days). In addition, here the reliability of the visual feedback was only manipulated at the task level and modulating contextual interference during the learning of a task (e.g., by manipulating the practice conditions) may facilitate even more the long-term learning process (for a review, see Wright et al., 2016). Future studies should manipulate reliability during practice in a within participants design.

The datasets presented in this study can be found in online repositories. The names of the repository/repositories and accession number(s) can be found at: https://osf.io/xknh4/.

The studies involving humans were approved by General data protection regulations (GDPR) of the University of Poitiers. The studies were conducted in accordance with the local legislation and institutional requirements. The participants provided their written informed consent to participate in this study.

CS and YB conceived and designed the research. GC, MB, and CS set up the experiment. MB performed the experiments. CS, MB, and YB analyzed the data. CS, YB, and MB interpreted the results of the experiments. MB and CS prepared the figures. MB, CS, and YB drafted the manuscript. CS, YB, GC, and MB edited and revised the manuscript. All authors contributed to the article and approved the submitted version.