Motor adaptation to environment changes predicting interactable object behaviour can be flexible and implicit

The human motor system can adapt to perturbations by updating existing models of motor control or by creating context-specific motor memories or strategies. We investigate if motor adaptation is context-informed when a perturbation is applied to either the throw direction of a ball, or the acceleration of the ball post-release during a virtual throw-to-target task. Using the visual slant of the task surface in an immersive virtual environment, we determine if the tendency for model updating is influenced by informative visual cues that predict perturbations to intended actions. Our findings reveal that perturbations resembling accelerations enabled flexible motor adaptation regardless of the presence of the slant cue. Perturbations in the throw direction conversely led to internal model updating. Additionally, visual slant properties of the task surface elicited implicit, slant-specific changes in performance. Our findings underscore the role of visual properties of both perturbations and environments in flexible motor learning.


Introduction
The ability to flexibly adapt to changing environments is crucial for successful performance in many real-world motor tasks.Upper limb motor adapta on has been extensively studied in the past by perturbing the hand movement or representa ons of the hand, leading to errors in predic ons of sensory consequences of movements and elici ng an upda ng of internal models governing ac ons.
Although such motor adapta on paradigms well represent internal changes to the body, such as injury or fa gue, a significant por on of motor learning in daily ac vi es involves the motor system adap ng to changes in objects and the surrounding environments.When adap ng to these externally localized perturba ons, there may be clear visual indicators that predict the behaviour of objects of interac on, such as the movement of grass on a golf course that predicts the strength of a sidewind.In such situa ons, rather than upda ng current model for the task, motor adap on may rely on forming contextspecific motor memories or cogni ve strategies, allowing for effec ve and rapid task switching when environments change.
Understanding how the brain adapts to real-world perturba ons is essen al for developing effec ve interven ons and training programs for individuals with motor impairments or those seeking to improve their motor skills.In this study, we tested if adapta on to perturba ons that affect the accelera on of a thrown object, sugges ng external sources of errors, leads to more flexible motor adapta on than the internal-model upda ng expected when adap ng to visuomotor rota ons, which suggest internal sources of error.Although par cipants adapted to both types of perturba ons, those adap ng to accelera ons in ball paths and even perturba ons visually resembling accelera ons did not show evidence for model upda ng.Instead, these perturba ons facilitated more flexible motor learning.
We addi onally tested if immersive visual cues that could plausibly explain errors or aid in strategyforma on could supress model upda ng.We find informa ve changes in the visual environment did not affect model upda ng characteris cs but could cue immediate and implicit changes in task performance during learning and de-adapta on.
In prior work, systematic differences between predicted and actual consequences of planned upper limb movements reliably lead to the updating of internal models of motor control that map goals to actions [1][2][3][4].Updating existing models avoids switching costs associated with motor memory selection and retrieval, allowing for efficient and timely movements [5,6].Motor errors can be reduced exponentially via one or more learning processes [7][8][9][10].However, this approach is not ideal when motor contexts change rapidly as these models must be re-updated to previously experienced contexts through the same learning processes, and errors must again be decreased exponentially over time.
More flexible motor learning, either through the creation of new implicit motor memories or through strategy formation, may improve task-success when movement types, objects of interaction or environments change rapidly [11].Although behaviour patterns indicating model updating have been observed when adapting to errors predicting a movement or hand-object interactions [12][13][14][15], the relative contributions of model-updating or more flexible motor learning when adapting to errors in predicting object-environment remains unclear.In this study, we use a target-directed ball throwing task to test if motor adaptation patterns differ when adapting to perturbations that are likely due to environment interactions, like accelerations added to the movement paths of thrown objects, or handobject interactions, like rotations of intended throwing directions.
Localizing the source of a given error during a motor task can be crucial for determining if existing internal models should be updated or if an alternative model or strategy should be created.
Perceptual cues the suggest a change in movement type, like changes in hand or body posture, reliably lead to flexible motor adaptation through the formation of new motor memories that can be switched to when returning to previously experienced conditions [16][17][18][19].When the movement type remains consistent, findings on the efficacy of interactable object or environment-based cues to facilitate flexible motor adaptation are mixed [20][21][22][23][24][25].Generally, properties extrinsic to movements, such as the colour of objects in the environment or object identity do not facilitate the creation of new motor memories [18,20,23].In some cases however, visual properties of the environment may serve to explain errors in outcomes under a given predictive model and render model updating less useful [26], allowing for flexible motor learning, possibly through the use of explicit strategies [12,15,24,25,27,28].Both saliency and task relevance are crucial for detecting changes in context and may explain these conflicting findings [29,30].Although cues like environment colour are salient, they may not facilitate context-change detection to the same degree as task-relevant cues [12,29,31].In dual-adaptation paradigms for example, although the colour of the target of reaching movements may fully indicate the presence of a target-colour-specific perturbation, the creation of target-colour-specific motor memories is not likely since the colour of external objects is usually irrelevant to internal models of upper limb movements.It is not clear however, if this holds when visible environmental changes fully predict changes in objectenvironment interactions.In studies where motor learning intrinsically involves learning the interaction between movement and the environment, properties of the object being interacted with are essential, as learning is often optimized to account for object or environment properties even when errors are low [32,33].Here, we examined whether adaptation patterns differ from those in upper-limb motor adaptation tasks when errors could be explained by object-environment-interactions in an immersive virtual environment.Specifically, we investigate if changes in internal model updating caused by our prescribed internally and externally attributed perturbation types is influenced by visual cues in the environment that reliably predict the behaviour of thrown objects, such as the slope of a surface on which thrown objects travel.
In a series of experiments, par cipants made realis c throwing movements to hit targets in a virtual task where the thrown object rolled on a flat surface.During the experiments par cipants adapted to two perturba on types: visuomotor rota ons at the moment of release, o en internally a ributed, and perturba ons resembling real-world accelera ons, ac ng on the object a er the release.
We test if accelera on-like perturba ons lead to flexible motor adapta on allowing for fast changes in performance in new or repeated motor contexts.We further test if immersive and informa ve visual cues, which may prompt faster motor adapta on, can also lead to more flexible motor learning for both internally and externally a ributed perturba on types.We find that accelera ons, and even perturba ons that are visually accelera on-like, can lead to more flexible motor learning than model upda ng.In all cases, immersive visual changes to the slant of the task surface prompted immediate behaviour change but did not facilitate the crea on of new motor memories or strategies required for flexible motor learning.In a subsequent experiment, these visual environment-cued behaviour changes persisted even while par cipants supressed cogni ve strategies during throwing movements.We conclude that behavior changes triggered by informa ve and task relevant changes in the environment can occur implicitly.This implies that the updated models in these instances include models of ballenvironment interac ons, rather than models specifically of the arm or release dynamics.

Apparatus/Setup
Par cipants sat on a height-adjustable chair with a table at waist height.Armrests were adjusted to match the table height and par cipants rested their elbows on the armrests for comfort between trials.Par cipants were then verbally instructed on the details of the task.The instruc ons used can be found at this project's Open Science Framework repository (osf.io/a5nv3/).A er receiving the verbal instruc ons, par cipants donned a head-mounted display system (HMD: Oculus Ri Consumer Version 1; resolu on 1080 by 1200 for each eye; refresh rate 90 Hz) and grasped Oculus Touch controllers in both hands.Three Oculus Constella on sensors tracked the posi ons of the HMD and Oculus Touch controllers.Par cipants received all visual feedback via the HMD.The virtual-reality environment was developed in Unity 3D, using the Unity Experiment Framework to handle trial and block schedules during the experiment [34].Before star ng the experiments, par cipants performed prac ce trials where they were encouraged to explore the tasks involved in the experiment.Par cipants could repeat the prac ce trials if required.Once all prac ce trials were completed, further instruc ons were provided by the experimenter and par cipants began the experiment.

Roll-to-Target Task
In all experiments, par cipants repeatedly performed the Roll-to-Target task.The goal of the task was to roll a ball such that it travelled as close as possible to a target loca on (Fig 1a).Par cipants received a score from 0-10 for each roll based on the minimum distance between the ball path and the target.Par cipants also received +5 bonus points for accurately hi ng the target.Each trial of the task began by par cipants holding their right hand near the midline of their body below the chin.In the virtual environment, they were situated over a large horizontal plane (4 m x 4 m) located at waist height, termed the Surface (Fig 1a).A wall with a horizontal wood plank or brick texture at the far end of the Surface, along with grid-lines along the task surface, provided an accurate visual representa on of the orienta on of the Surface.A virtual 10-cm diameter ball was placed in front of the par cipant at the "star ng posi on".A target was placed 1 meter away from the star ng posi on of the ball in one of 4 possible loca ons on the Surface (84°, 88°, 92° and 96° in polar coordinates: Fig 1a).Target loca ons were pseudorandomized so that in every 4-trial set, par cipants threw to each target at least once.
The goal of the Roll-to-Target task was to make a throwing movement to roll a ball and hit targets as accurately as possible.Each trial contained an Ac on step and a Feedback step.Each trial began with an Ac on step, where par cipants made a throwing movement to the target by holding down the trigger bu on on the Oculus Touch controller with their index finger and making a quick, outward mo on with their dominant hand.The ball was released when the hand moved 15 cm from the star ng posi on of the throw, or when the trigger bu on was released, whichever event occurred first.
To ensure consistent rolling mo ons, par cipants needed to complete the Ac on step within 1.5 seconds from the start of the trial to receive a score for the trial.Although par cipants did not receive a score under this me-out condi on, we included these trials in our analysis.The Feedback step began immediately upon the release of the ball.
During the Feedback step, par cipants observed the ball roll and were given feedback on the ball path.The speed and direc on of the throwing movement at the point of release in the Ac on step determined the speed and travel path of the ball in the Feedback step.The speed of the ball was capped at a maximum of 3.96 m/s.The feedback step ended either a er hi ng the target or a er 2 seconds had elapsed from the start of the Feedback step.At the end of the Feedback step par cipants were shown a visual representa on of the ball path and received a score for the trial based on the minimum distance between the ball path and the target (Fig 1b-c).

Task Varia ons
We used 4 varia ons of the Roll-To-Target task that modified the rela onship between the throwing movement in the Ac on step and the ball path in the Feedback step: Aligned, Rotated, Accelerated, and Curved (Fig 1d).While all par cipants performed Aligned Roll-To-Target tasks, the Rotated, Accelerated, and Curved varia ons were considered perturba ons and each par cipant only experienced one of the 3, dependent on their group (see Experiment Protocol).
In Aligned varia ons of the task, the visual Surface was aligned to the real-world floor-plane.In all other varia ons, the Surface was either in the same visual alignment as the Aligned varia on, or displayed with a 25° CCW visual slant (Fig 1c) in groups termed "Uncued" and "Cued" respec vely (Fig 1d-e).
In Aligned Roll-to-Target tasks, the direc on and speed of the throw in the Ac on step determined the ini al heading direc on and speed of the ball in the Feedback step.In Aligned tasks, we con nuously applied 'Baseline Accelera ons' to the ball path (-0.31, -0.15, 0.15, 0.31 m/s 2 , pseudorandomized).Baseline Accelera ons were used so that all Roll-to-Target trials involved changes in the ball's direc on of travel.Baseline accelera ons were small, centered around zero and had an unpredictable order to have li le to no effect on task performance.
In Accelerated Roll-to-Target tasks, the ini al direc on of the ball movement during the Feedback step matched the throw direc on in the Ac on step, while a constant le ward accelera on of 2.49 m/s 2 was applied to the ball as it rolled during the Feedback step (Fig 1f).The movement path of the ball during this task varia on resembled a plausible movement path on a surface with a 25° surface slant.Although all Roll-to-Target tasks included accelera ons in the ball path during the Feedback step, only the accelera on in Accelerated tasks could be reliably countered.
In Rotated Roll-to-Target tasks, the speed of the throw in the Ac on step determined the ini al speed of the ball, but the ini al heading direc on of the ball in the Feedback step was rotated by 30° or 15° counter-clockwise rela ve to the throw direc on along the Surface plane (Fig 1f).We applied Baseline Accelera ons to all Rotated Roll-to-Target tasks.We used a rota on size of 30° and an accelera on of 2.49 m/s 2 in the experimental groups to match the changes in throw angles required to reliably hit targets with Rotated and Accelerated perturba ons.In a follow up experiment, we used a rota on size of 15° to match the error sizes par cipants observed in the Accelerated perturba ons.
In Curved Roll-to-Target tasks, the ini al speed and direc on of the ball movement during the Feedback step matched the throw speed and direc on during the Ac on step.During this task varia on, the path of the ball curled in a counter-clockwise direc on so that the ball passed through a point 30° counter-clockwise to the throw direc on when at the target distance (Fig 1f).The magnitude of the angular offset of the ball path was scaled to the distance from the star ng posi on of the ball.Curved perturba ons required the par cipants to match the throw speeds and throw angles required to reliably hit targets in the 30° Rotated Roll-to-Target tasks.A er comple ng prac ce trials, an experimenter reiterated the task objec ves.These instruc ons were also displayed within the virtual reality environment and par cipants were encouraged to mentally follow along (instruc ons found at osf.io/a5nv3/).The experiments were divided into mul ple phases (Fig 1g).All experiments began with a Baseline phase where par cipants performed 40 Aligned Roll-to-Target tasks.These tasks were performed in trial-sets of 4 trials within which par cipants rolled the ball to each of the 4 possible targets and experienced each of the 4 possible Baseline Accelera ons.The order of the target loca ons and Baseline Accelera ons were randomized independently of each other within the trial-sets.The visual task surface was always aligned with the floor plane during the Baseline phase.Par cipants in the Cued and Uncued Aligned and Rotated groups performed an addi onal 40-trial Baseline phase using their non-dominant hands.This data was not analyzed for this study but is included in the data repository (osf.io/a5nv3/).Following the Baseline phase, par cipants were asked to take a short break for a minimum of 1 minute.The experimenter then read further instruc ons informing the par cipants that although the environment may change, their goal was to always hit the target with the ball as accurately as possible.

Experiment Protocol
Par cipants in all groups then completed the Training phase (80 trials) where they performed one of 4 varia ons of the Roll-to-Target tasks, determined by their assigned group.The visual task surface was slanted by 25° (CCW roll) in all Cued groups and was aligned with the real-world floor plane in all Uncued groups.During the Training phase, the colour of the task surface was changed from black to green in the Cued groups in the four ini al experiment groups, while the colour of the surface was changed to green for all groups in the follow-up groups.
To test the decay of learned behaviour when returning to baseline-like condi ons, par cipants

Data Analysis
For each trial, we calculated the angle of the throwing mo on in the Ac on step rela ve to a straight-line to the target along the task surface plane (termed "Throw Angle"), the speed of the throw, and the minimum distance to the target along the ball's travel path.The Throw Angle was our primary measure of interest as par cipants were required to modify it to successfully hit targets in all our perturbed condi ons.During Rotated and Curved Roll-to-Target tasks, changes in Throw Angles alone determined task success, while the speed of the throw also impacted task success in Accelerated Roll-to-Target tasks.We also calculated Minimum Errors for each trial: the smallest distance from any point on the ball path during the Feedback step to the centre of the target.
To account for individual differences in baseline throwing characteris cs, we calculated the median Throw Angles of each par cipant when throwing to each of the 4 possible target loca ons during the Baseline Phase.We subtracted these target-specific baseline Throw Angles from the calculated Throw Angles in the Training and Washout phases.We similarly determined baseline errors for each target by calcula ng the median magnitude of error experienced during the Feedback step of the Roll-to-Target task.To obtain a measure of addi onal error due to our imposed perturba ons, we subtracted these target-specific errors from the errors experienced during the Training and Washout phase.We analysed the fit parameters for the start points and decay rates during the Washout phase (Eq 2) to determine if adapta on to our imposed perturba ons led to the upda ng of exis ng models of motor control or the crea on of new models.In instances of model-crea on, we expected performance to return to baseline condi ons quickly once a context change was detected during the Washout phase, signalled by low ini al start points or fast decay rates.Alterna vely, in instances of model upda ng, we expected high start points and rela vely slower decay of learned behaviour during the Washout phase.
In Experiment 1, we tested whether a visual environmental cue or the perturba on type affected if internal models of motor control were updated.We used a 2x2 between-subject design where each par cipant experienced one of the 4 possible permuta ons of Visual Cue (Cued or Uncued) and Perturba on type (Accelerated or Rotated) combina ons during the Training phase.
In Experiment 1, we first determined if the presence of the visual cue or perturba on type affected the characteris cs of par cipants' learning curves.We compared learning rates and asymptotes in the Training phase by conduc ng two 2x2 between-subject Analysis of Variance (ANOVA) tests with the presence of the visual cue and perturba on type as between-subject factors.When effects were found we also calculated and reported an effect size (generalized η 2 ) and computed and reported inclusion bayes factors between models that included and did not include relevant effects [35].We used Tukey's HSD for post hoc pairwise analyses when necessary.
To determine if par cipants could flexibly change motor behaviour when experiencing previously seen contexts, we then compared start points and decay rates during the Washout phase with two addi onal 2x2 between-subject ANOVAs with the presence of the visual cue and perturba on type as between-subject factors.We again calculated generalized η 2 , inclusion bayes factors and Tukey's HSDs when effects were found.
We repeated these 4 2x2 between-subject ANOVAs for Follow-up Experiment 1 and 2. In Followup Experiment 1, we compared Cued and Uncued Rotated-15 groups with the original Cued and Uncued Accelerated groups.In Follow-up Experiment 2, we compared the original Cued and Uncued Rotated groups with Cued and Uncued Curved groups.
In Follow-up experiment 1, we conducted addi onal 2x2 ANOVAs, with visual cue and perturba on type as between subject factors, comparing start points and error reduc on rates fit to the reduc on in Minimum Error during the Training Phase.We conducted two sets of 2x2 ANOVAs, first comparing the start points and error reduc on rates of Minimum Error in the Accelerated groups to those in the Rotated groups, and second comparing the start points and decay rates of Minimum Error in the Accelerated groups to those in the Rotated-15 groups.
In Experiment 2, we tested whether environmental changes affected model upda ng.To elicit the upda ng of exis ng internal models, the Training phase was iden cal to the Cued Rotated group.We fit exponen al decay func ons (Eq 2) to the decay of learned behaviour during the 8-trial test phases.To determine if environment changes led to implicit performance changes, we then compared changes in Throw Angles during Test blocks following 180° pivots against those a er 360° pivots using two paired ttests.For all tests, we reported Bayes factors and Cohen's d when effects were found.

Results
Experiment 1: Effects of visual cues and perturba on types on context-based flexibility of motor learning.Although weak, we controlled for this effect in a follow-up experiment using Curved perturba ons (see Follow-up experiment 2).Overall, the slant of the task surface acted as an informa ve visual cue when adap ng to both Accelerated and Rotated perturba ons, allowing for more rapid motor adapta on.The surface slant cue affected both the Start Points and Decay Rates during the washout of learned behaviour.To determine if this cue served to ini ate a context switch, we reanalyzed washout dynamics when including plausible context-switch detec on in all groups.That is, we assumed that for the "Cued" groups, the change in surface slant during the first trial of the Washout phase would be sufficient to cue a context change, whereas, for the "Uncued" groups, the visual errors experienced during the first trial of the Washout condi on would be necessary.To be er compare the two groups, we included performance during the last trial of the Training phase for the "Cued" groups, ensuring our modelled Washout behaviour included context changes in all groups.We define Trials+ to be the relabelled trial numbers used for plo ng and model-fi ng.For all subsequent analyses, we used Trials+ to analyse the decay of motor adapta on during the Washout phase.Analyses of the uncorrected decay behaviour can be found in this project's Open Science Framework repository (osf.io/a5nv3/).
When ).The decay of adapta on to accelerated perturba ons was fast, taking on average 2 trials to return to baseline-like performance, sugges ng motor learning that is flexible when context changes are detected.Following adapta on to rotated perturba ons, the lower decay rates suggest exis ng internal models were updated.
For each trial, we defined Minimum Errors as the shortest distance between any point on the A er confirming that error reduc on rates and star ng errors were similar across all groups, we repeated the analyses in Experiment 1 using normalized Throw Angles.suggests that the differences in the propensity for model upda ng was not influenced by the observed differences in error sizes, but the types of the perturba on experienced.
When adap ng to Accelerated perturba ons, par cipants could hit the target with the thrown ball through specific combina ons of Throw Angles and Throw Speeds.

Discussion
In our virtual ball-throwing task, the tendency for model upda ng during motor adapta on was en rely determined by the type of perturba on experienced.When errors could be plausibly a ributed to object-environment interac ons, as in real-world-like accelera ons of ball paths, motor learning was flexible, and performance could quickly switch between recently learned and previously experienced contexts.When the visual feedback of the hand-ball interac ons at the moment of release were misaligned instead, sugges ng internal sources of error, we observed evidence for the upda ng of internal models of motor control.Immersive visual cues about the task environment that reliably predicted motor-context switches enabled faster ini al learning and fast switching to exis ng internal models but did not facilitate the crea on of new motor memories.Although changes in the immersive virtual environment during perturba ons did not determine if models were upda ng during motor learning, proper es of the virtual environment, like the surface slant, did inform the implicit internal models involved in execu ng a target-directed ball-rolling ac on.
In our study, Accelerated perturba ons led to motor adapta on that could be flexibly changed when previously experienced contexts were detected.Accelerated perturba ons can plausibly be explained by physical processes ac ng on the rolling ball, such as the effects of gravity when rolling on a slanted surface or the effects of sidewinds when rolling on a flat surface.Par cipants may have intui vely simulated physics proper es that govern interac ons in the environment [36,37] to explain away unpredicted behavior and supress model upda ng [26].Alterna vely, recent findings suggest that people fail to accurately es mate accelera ons while s ll compensa ng for accelera on-based errors over me [38].It would then follow that accelera on-like errors, like the Curved perturba ons in our follow up experiment that could not be explained by Newtonian mechanics, should also lead to learning similar to accelera on based perturba ons.That is indeed what we observed.Surprisingly, visual explana ons for the accelera on perturba on, that is, the slant of the surface, did not have any addi onal effects on the tendency for model upda ng, sugges ng visual informa on did not provide any addi onal informa on for the assignment of error to an external source.
Although motor changes during adapta on were matched across Accelerated and Rotated groups, par cipants in the Rotated groups experienced larger errors during trials in the early Training phase.Larger errors may also lead to slower decay, both due to higher amounts of implicit adapta on, and models being re-updated taking longer to return to baseline.Learning rates should be similar across learning rates when errors are similarly reliable [8][9][10], but large error sizes could lead to large explicit changes in performance [39].Findings in our follow-up study using 15° visuomotor rota ons suggest the perturba on type, and not experienced error sizes, determined if internal models were updated.
Although experienced errors in the 30° Rotated group were large, they may not have been large enough to elicit the use of large strategies that could be suppressed or changed in response to a changes in the external environment [39,40].
A successful real-world, target-directed throw can be achieved by various combina ons of throwing speeds and throwing direc ons due to redundancies in the task.The solu on subspace for our Accelerated tasks, i.e., a manifold along the solu on space that led to successful task outcomes, allowed for higher variability in throwing direc ons than the Rotated condi ons.When task solu ons allow for redundancy, variability is only reduced to a limited degree [41,42].A wide solu on subspace may allow for both increased motor explora on along the throwing angle dimension as well as reduced reinforcement learning to throwing direc ons specific to the perturba on, allowing for faster return to baseline-like performance [43][44][45].However, since adapta on to Curved perturba ons where the solu on space was iden cal to the Rotated perturba ons was flexible, our findings suggest that visually observed proper es of the perturba on, and not the solu on space determine if models are updated.
In our follow-up experiments, although the colour of the task surface changed during the training phase, only the visual change in the task-surface slant reliably cued an immediate change in motor performance.Our findings support previous work sugges ng colour changes in the environment are ineffec ve at elici ng dual learning, where the a crea on of mul ple context specific motor memories is necessary [18,20,27].Generally, proper es of the task environment that do not directly affect movement dynamics, termed "indirect cues", do not reliably lead to the crea on of new contextspecific motor memories [18,20,23].In our study however, both the visual slant and colour of the task surface were indirect cues containing no direct informa on about dynamic body states during a throwing movement.Our findings from Experiment 1 and Experiment 2 suggest that environment proper es serving as indirect cues that explain the behaviour of interactable objects, as opposed to less taskrelevant cues, are effec ve at promp ng the development of both explicit and implicit changes in behaviour.
Although prior works suggest likely strong contribu ons of explicit learning processes to visual slant-related learning [27], in this study, we do not dis nguish whether a lack of model upda ng was due to the crea on of cogni ve strategies to counter the perturba on, or the crea on of new implicit motor memories.Indeed, neither the use of explicit strategies during motor learning nor implicit motor learning processes prevent the parallel development of either class of adapta on [39,40,46,47].In Experiment 2, we directly tested for implicit components of surface-slant specific adapta on.When par cipants were asked to exclude any strategies while throwing to targets on a slanted surface, we observed implicit slant-specific motor learning.Since we relied on verbal instruc ons, there is a chance that par cipants did not disengage explicit slant-specific aiming strategies.However, in such cases, we expect throws to the opposite direc on when the task surface was pivoted 180°.Instead, all but one par cipant reduced their trained throwing angles in response to 180° pivots while maintaining the learned devia on direc on, sugges ng that the observed slant-specific behaviour changes were indeed implicit.
Overall, our results suggest that adapta on to accelera on-like perturba ons to the path of a thrown object is flexible and does not rely on upda ng internal models of motor control.Addi onally, informa ve task relevant environment cues may facilitate immediate and implicit behaviour changes but do not affect the propensity for internal model upda ng during motor learning.
Par cipants 169 par cipants (103 female, 20.75 ± 4.55 years of age, mean ± SD) par cipated in the study and were divided into one of 8 groups within 4 experiments.All par cipants had normal or corrected-tonormal vision and were naïve to visuomotor learning experiments.All par cipants voluntarily par cipated in the study and provided wri en and informed consent.The procedures used in this study were approved by York University's Human Par cipant Review Commi ee and all experiments were performed in accordance with ins tu onal and interna onal guidelines.

Figure 1 :
Figure 1: a: Visual display in HMD at the start of each trial in the Roll-to-Target task.The target appeared at one of 4 possible target loca ons.b: Visual display at the end of the Feedback phase of a Roll-To-Target task with the task surface aligned with the real-world floor and Baseline-accelera ons.c: Visual display at the end of the Feedback phase of a Cued-Curved Roll-to-Target task.d: Possible perturba on/cue combina ons during the Training and Transfer phases of Experiment 1. e: Possible immersive visual slants of the task surface.During Cued Roll-to-Target tasks in Experiment 1, the task surface was slanted 25°.In Experiment 2, a 25° visual slant acted as the cue during the Training phase and the anima on prior to each Washout phase resulted in a -25° or 25° visual slant for 180° and 360° anima ons respec vely.f: Possible perturba ons during the Feedback step of Roll-To-Target tasks.g: Phases and trials in Experiment 1. Surface Baseline* and Transfer* phases were not analyzed in this study.Par cipants used their non-dominant hands during Baseline* and Transfer* phases.h: Phases and trials in Experiment 2.
completed a Washout phase consis ng of 40 trials of the Roll-to-Target task immediately following the Training phase.For all groups, the visual task surface was aligned with the floor plane during the Washout phase.Par cipants in the Cued and Uncued Aligned and Rotated groups performed an addi onal 40trial Transfer phase with their non-dominant hands.The Transfer phase was iden cal to the Training phase.This data was not analyzed for this study but is included in the data repository (osf.io/a5nv3/).

Experiment 2 :
Effects of visual surface-slant cues on implicit motor learning.
In a second experiment (1 group: n=20, 14 female, 22.40 ± 5.88 years of age), we tested whether internal model-upda ng during motor adapta on relies on visual environmental cues -namely the visual slant of the surface in the Roll-To-Target tasks.In this experiment we introduced Pivots: anima ons of the virtual task-surface played between phases of the experiment (Fig 1e, right side).Each Pivot lasted 2 seconds and involved a counterclockwise (CCW) rota on of task surface on the yaw axis, either by 180° or 360°.During the Pivot anima ons, par cipants were verbally instructed via a recording played in the head mounted display system to disengage any cogni ve strategies they may have used during the Training phase.Pivots were always followed by 8 Aligned Roll-to-Target trials to test the effect of the Pivot on performance (Fig 1h).Par cipants first completed a Baseline phase, performing 2 sets Aligned Roll-to-Target tasks (16 trials, then 8 trials) separated by a 180° Pivot.Par cipants then repeated the 2 sets of Aligned Roll-to-Target task, separated by a 360° Pivot.During the Training phase, par cipants performed 80 Rotated Roll-to-Target trials with a rota on magnitude of 30°, and the visual task surface was slanted by 25° (CCW roll), as in the Cued groups in Experiment 1.The colour of the task surface was changed to green during the Training phase.During Test phases, we used 8-trial blocks to test the effect of a 180° or a 360° pivot of the task surface.We repeated the test blocks so that par cipants performed 4 test blocks following a 180° pivot, and 4 test blocks following a 360° pivot.Par cipants performed top-up Training blocks (20 trials of the Cued Rotated Roll-to-Target task) in between each set of test blocks (Fig 1h).

1 ) 2 )
To compare across groups, we normalized each par cipant's Throw Angles rela ve to their median Throw Angles during the middle of the Training phase (trials 33-40).We then fit exponen al decay func ons to the change in each par cipant's normalized Throw Angles during the first 40 trials of the Training and Washout phases.plearn = a -a(1 -r) t-1 (Equa on 1 represents the change in performance over trials (t) in the Training phase (plearn) given a par cipant's learning (r) and the asymptote of adapta on (a).pdecay = s(1 -r) t-1 (Equa on 2 represents the decay in learned behaviour over trials (t) in the Washout phase (pdecay) given a par cipant's start point of decay (s) and decay rate (r).We assumed all par cipants started at baseline-like performance at the start of the Training Phase and returned to baseline-like performance at the end of the Washout phase.To visualize the reduc on in errors over me during the Training phase, we also fit the exponen al decay formula (Eq 2) to the changes in Minimum Errors over me in the Training phase.

Figure 2 :
Figure 2: a: Throw angle devia ons from straight-to-target throws ("Throw Angle"), rela ve to individual baseline Throw Angles, over the Training and Washout phases in Experiment 1. b: Throw Angles during the first 40 trials of Training and Washout phases.These trials were used for fi ng exponen al decay models.c: Normalized Throw Angles during the first 40 trials of the Training phase.d: Exponen al decay func ons fit to the changes in Normalized Throw Angles over trials in the Training phase.e: Learning rates and asymptotes of adapta on of fit Normalized Throw Angles of par cipants.f: Normalized Throw Angles during the Washout phase.g: Exponen al decay func ons fit to the changes in Normalized Throw Angles over trials in the Washout phase.h: Start points of decay and decay rates of Normalized Throw Angles of par cipants in the Washout phase.i: Normalized Throw Angles during the Washout phase including trials prior to cue detec on (Trials+) in all Cued groups.j: Exponen al decay func ons fit to the changes in Normalized Throw Angles over Trials+ in the Washout phase k: Start points of decay and decay corrected for the detec on of context changes, only the perturba on type affected the Start Points of decay func ons fit to washout behaviour (main effect: F(1, 76) = 4.76, p = 0.032, η 2 = 0.059, BFincl = 1.34;Fig 2i-k).Accelerated perturba ons led to lower Start Points of decay (meanstart points (accelerated) = 0.815) than Rotated perturba ons (meanstart points (rotated) = 0.981, Tukey HSD = 0.0323, d = -0.488;Fig 2k: le plot).Differences in the solu on space between the perturba on types may have led to the differences in the Start points of decay.We explore this possibility in Follow-up Experiment 2. Perturba on type alone determined the decay rates of learned devia ons of Throw Angle (main effect of perturba on type: F(1, 76) = 75.59,p < .001,η 2 = 0.499, BFincl = 1.49*10 10 ; Fig 2k: right plot).Adapta on to accelerated perturba ons decayed more rapidly than adapta on to rotated perturba ons (meandecay rate (accelerated) = 0.632, meandecay rate (rotated) = 0.219; Tukey HSD < 0.001, d = 1.95.
ball's travel-path during the Feedback phase and centre of the target.Although we designed our task to match the adapta on rates of Throw Angles during the Training phase for both Accelerated and Rotated groups, par cipants in the Accelerated groups experienced smaller Minimum Errors during the Training phase (Fig 3a-b).When fi ng Eq 2 to Minimum Errors experienced in the Training phase, par cipants in the Accelerated groups experienced lower ini al Minimum Errors (main effect of perturba on on star ng points: F(1, 76) = 57.00,p < 0.001, η 2 = 0.429, BFincl = 9.52*10 7 ; meanini al minimum error(accelerated) = 1.67, meanini al minimum error(rotated) = 3.37; Tukey HSD < 0.001, d = -1.69)and reduced Minimum Error more quickly (main effect of perturba on on learning rate: F(1, 76) = 7.05, p = 0.010, η 2 = 0.085, BFincl = 3.75; meandecay rate(accelerated) = 0.244, meandecay rate(rotated) = 0.125; Tukey HSD = 0.010, d = 0.594; Fig 3b-c).To determine if differences in Minimum Error reduc on led to the choice between model upda ng or flexible motor learning, we conducted a follow-up experiment matching the Minimum Error experienced between the two perturba on types.To do so, we repeated the Cued and Uncued Rotated condi ons with a second pair of Rotated condi ons using a 15° rota on to the Throw Angle: the "Rotated-15" groups (Fig 1d-e: dark red and dark yellow).To ensure both the Accelerated groups and the Rotated-15 groups experienced a similar evolu on of Minimum Errors over me, we again fit Eq 2 to Minimum Error over the first 20 trials of the Training phase (Fig 3d-e).When adap ng to 15° rota ons, ini al Minimum Errors were more likely under models that did not include effects of perturba on type, cue, or interac ons between the two (BFincl = 0.258, 0.539, 0.147, respec vely; Fig 3f: le plot).Similarly, the rates of Minimum Error reduc on between the Accelerated groups and the Rotated-15 groups were more likely under models that did not include effects of perturba on type, cue, or interac ons between the two (BFincl = 0.190, 0.265, 0.076, respec vely; Fig 3f: right plot).

Figure 3 :
Figure 3: Minimum Errors during Training phases and Throw Angles during Training and Washout phases.a: Error experienced during the Training phase in Accelerated and Rotated groups.b: Exponen al decay func ons fit to the reduc on of error over trials in the Training phase in Accelerated and Rotated groups.c: Start points and decay rates of fit Minimum Errors experienced by par cipants in Accelerated and Rotated groups.d: Error experienced during the Training phase in Accelerated and Rotated-15 groups.e: Exponen al decay func ons fit to the reduc on of experienced error over trials in the Training phase in Accelerated and Rotated-15 groups.f: Start points and decay rates of fit Minimum Errors experienced by par cipants in Accelerated and Rotated-15 groups.g: Normalized Throw Angles during the Training phase in Accelerated and Rotated-15 groups.h: Exponen al decay func ons fit to the changes in Normalized Throw Angles over trials in the Training phase in Accelerated and Rotated-15 groups.i: Learning rates and asymptotes of adapta on of fit Normalized Throw Angles of par cipants in Accelerated and Rotated-15 groups.j: Normalized Throw Angles during Trials+ in the Washout phase in Accelerated and Rotated-15 groups.k: Exponen al decay func ons fit to the changes in Normalized Throw Angles over Trials+ in the Washout phase in Accelerated and Figure4ashows the solu on space, the combina ons of Throw Angle and Throw Speed used during the Training phase of the experiments and their associated Minimum Error sizes, for each target when adap ng to Accelerated perturba ons.The solu on manifold is the sub-space within the solu on space that led to direct target hits, indicated by white-coloured data points.The solu on manifolds for Accelerated perturba ons were a func on of both Throw Angle and Throw Speed (Fig4a), whereas the solu on manifolds for Rotated perturba ons was a func on of only the Throw Angle, with some added noise due to the Baseline Accelera ons (Fig4b).Addi onally, a larger range of Throw Angles could lead to task success when adap ng to Accelerated perturba ons (Fig4a).To determine if these solu on space differences during the Training phase affected our findings, we conducted a second follow-up experiment where the ballpath during the Feedback step of the Roll-to-Target task was curved (see Fig1f).The ball path was deviated from a straight-to-target path by an amount propor onal to the distance from the star ng posi on, so that when at the target distance, the angular devia on rela ve to the star ng posi on is always 30°.Curved perturba ons resulted in solu on manifolds like Rotated perturba ons (Fig4b-c) while retaining visual feedback proper es of Accelerated perturba ons (Fig1f).Once we ensured similar solu on manifolds for both perturba on types, we compared changes in Throw Angles in the Curved groups with those in the Rotated groups to determine if our findings from Experiment 1 were driven by the different solu on manifolds.

Figure 4 :
Figure 4: Minimum error sizes experienced (where white data points indicate a direct target hit) during throws in the Feedback step during the Training phase in Accelerated (a), Rotated (b), and Curved (c) groups as a func on of Throw Angles and throw speeds in the Ac on step.d: Normalized Throw Angles during the Training phase in Curved and Rotated groups.e: Exponen al decay func ons fit to the changes in Normalized Throw Angles over trials in the Training phase in Curved and Rotated groups.f:

Figure 5 :
Figure 5: a: Changes in Throw Angles over 8-trial Test phases in the Pivot Experiment, along with the first 8 trials in the Washout phase of the Cued and Uncued Rotated 30° groups from Experiment 1. b. Start points of decay and decay rates during 8-trial Pivot 180 and Pivot 360 Washout phases, along with start points of decay and decay rates during the first 8 trials of the Washout phase of the Cued and Uncued Rotated 30° groups from Experiment 1. Shaded areas represent 95% confidence intervals.