Abstract
Humans rapidly update the control of an ongoing movement following changes in contextual parameters. This involves adjusting the controller to exploit redundancy in the movement goal, such as when reaching for a narrow or wide target, and adapting to dynamic changes such as velocity-dependent force fields (FFs). Although flexible control and motor adaptation are computationally distinct, the fact that both unfold within the same movement suggests that they interact functionally to support task-specific adjustments. To test this hypothesis, we conducted a series of experiments combining changes in the target structure and a force field presented separately or in combination. Seventy-six human participants (both sexes) took part in this study, with each experiment involving different participants. They were asked to reach for a target that could change from a narrow square to a wide rectangle between or during trials. Step loads were used to assess whether participants exploited target redundancy. In a separate experiment, we added a force field in addition to target changes and step loads. Our results revealed a reduced ability to exploit target redundancy when sudden target changes occurred concurrently with FF adaptation. Furthermore, the magnitude of adaptation was reduced when step loads were added to the FF. Crucially, this interference emerged specifically when all perturbations impacted motor execution simultaneously. These results indicate that flexible control and motor adaptation interact in a nontrivial manner, suggesting interdependence between these behavioral mechanisms, and a clear identification of the timescale at which they are engaged—namely, during movement.
Significance Statement
Humans rapidly adapt to changes in task demands, such as target structure changes or exposure to force fields (FFs). These two types of adjustments occur within a single movement, suggesting potential interactions between them. Our experiments revealed that the combination of FF exposure with online target shape changes selectively reduced participants’ ability to exploit target redundancy, while the combination of FF and step loads led to a reduced extent of motor adaptation. These findings confirm that motor adaptation occurs not only between trials but also during movement. The selective nature of the observed interference highlights an interplay between flexible control and motor adaptation, underscoring the importance of understanding the timing of these processes to better characterize their underlying neural circuits.
Introduction
Reaching movements are fundamental daily actions that sometimes require flexibility to cope with external perturbations like forces applied to the hand or changes in target shape or position. These factors impact planning and execution strategies, as evidenced by the effects of visual and mechanical perturbations on movement kinematics and feedback control strategies (Franklin and Wolpert, 2008; Izawa and Shadmehr, 2008; Kurtzer et al., 2008; Crevecoeur et al., 2013).
Several studies have explored humans’ ability to select control policies based on task demands by designing experiments where participants have flexibility associated with a state variable (position or velocity). For instance, humans select different control policies for shooting or stopping at a target (Liu and Todorov, 2007; Česonis and Franklin, 2022). Similarly, reaching behavior differs depending on target shape, with greater endpoint dispersion for wide bars than dots (Nashed et al., 2012). This supports the minimal intervention principle (Todorov and Jordan, 2002a), which states that deviations from trajectory are corrected only if they interfere with the task goals. De Comite et al. (2021) further showed that such flexibility in corrections occurred even when the target changed during movement, highlighting control policy updates during movement, including adjustments to obstacles. Similar flexibility was reported by Crowe et al. (2023), who demonstrated that online corrections also adapt to environmental changes.
In parallel, humans also adapt to new dynamic environments to recover the intended performance when exposed to a predictable perturbation. For instance, under a force field (FF) that deviates the hand from the straight path, participants gradually reduce the deviation, restoring straighter movements (Shadmehr and Mussa-Ivaldi, 1994; Dizio and Lackner, 1995; Singh and Scott, 2003; Lackner and DiZio, 2005). Feedback mechanisms play a central role in motor adaptation, with the brain adjusting a task-specific feedback controller to the requirements of novel motor skills (Ahmadi-Pajouh et al., 2012; Cluff and Scott, 2013; Joiner et al., 2017; Maeda et al., 2020). In addition to controller adjustments between movements, a fast timescale for motor adaptation is also present, influencing the execution of a single movement when different FFs are presented randomly to the participants (Crevecoeur et al., 2020b).
Much of the experimental work about reaching movements has been guided by optimal feedback control theory, which posits that task goals and environmental dynamics shape efficient movement strategies. In this framework, an optimal goal-dependent control strategy determines feedback gains during movement based on task constraints and limb dynamics (Todorov and Jordan, 2002b; Scott, 2004; Liu and Todorov, 2007). We focus on two types of control changes: those related to task demands and those related to changes in dynamics. Mathematically, task demands are expressed through the cost function, representing the goals and constraints of the task. For example, when the goal target switches during movement, penalties on behavioral performance are affected. In contrast, changes in dynamics refer to adjustments in internal models representing limb or environment dynamics. Although these operations are theoretically distinct, both changes in movement due to task updates and adaptation to changes in dynamics occur within ∼250 ms following a contextual perturbation (Kalidindi and Crevecoeur, 2023). Considering a typical movement time of ∼500 ms, this similarity in timing suggests that both controller updates and adaptation may occur within a reaching movement.
To explore their interdependence, we examined behavioral interference during a motor task with changes in goal target structure and step loads (“flexible control”) with or without FF (“adaptation”). Two hypotheses were formulated: if participants performed similarly in separated and combined subtasks, it would suggest independent mechanisms; conversely, behavioral differences would indicate interference. Consistent with the latter hypothesis, we found differences in participants’ ability to exploit target redundancy when exposed to the FF, but only when the target switch occurred during movement. Moreover, motor adaptation was reduced when step loads were applied during movement. Importantly, these interferences were not uniform across all conditions but emerged selectively when both adaptive and flexible control mechanisms were recruited simultaneously. These results indicate the interdependence between online modification of the control policy and motor adaptation, suggesting that both processes may rely on a functionally related mechanism within the motor system.
Materials and Methods
Participants
This study involved three groups of participants. The first group of 20 participants (18 right-handed and 2 left-handed, 11 females), aged from 20 to 31 was assigned to main Experiment 1. The second group, composed of 20 right-handed participants (11 females) with ages ranging from 21 to 26, performed main Experiment 2. Note that these two experiments involved different participants; hence, we chose to refer to these datasets as two different experiments. For the control experiments, 36 participants (28 females) were involved, with ages ranging from 18 to 32, and divided into three groups of 12 participants, one for each control experiment. All participants were unaware of the purpose of the study, had normal or corrected to normal vision, and had no known neurological or motor disorder. The procedures were approved by the ethics committee at the host institution (Comité d’Éthique Hospitalo-Facultaire, UCLouvain) and participants received financial compensation for their time.
Setup
All experiments were conducted using the KINARM endpoint robotic device. Participants were seated on an adjustable chair, facing the robotic device. They were instructed to grasp the handle of the right robotic arm using their right arm. This robotic arm allowed movements in the horizontal plane. They were seated at the beginning of each trial in such a way that their elbow and shoulder lay in a plane orthogonal to the movement plane, as can be seen in Figure 1A. The direct view of both the participant's hand and the robotic arm was blocked. An augmented-reality display was used to project the hand-aligned cursor (radius 0.5 cm), the virtual targets (Fig. 1A), and a score corresponding to the sum of all successful trials to keep participants engaged in the task. Muscular activity of two muscles of interest (pectoralis major and posterior deltoid) was measured using surface electrodes (Bagnoli surface EMG Sensor, Delsys). These muscles were selected because, given the arm configuration used in this study and based on previous studies (De Comite et al., 2021), they are strongly recruited to counter rightward and leftward step loads.
Experimental design and statistical analysis
For all the experiments, each trial followed the same sequence. At the beginning of the trial, the home and goal targets were displayed in gray. Participants had to move the hand-aligned cursor to the home target, which turned green when participants reached it. The goal target became blue when participants were allowed to initiate their movement. This go-cue was triggered after a random delay uniformly distributed between 2 and 4 s after reaching the home target. After this go-cue, participants could initiate their movement whenever they wanted so there were no constraints on the reaction time. The trial was successful if participants reached the goal target between 250 and 550 ms after they left the home target and stabilized in it for at least 500 ms. If the trial was successful, the goal target turned green, and the score projected onto the workspace was incremented by 1 point. If the movement was too slow or fast (movement duration outside the time window), the goal target turned red or black with a red outline, respectively. If the target was missed, it turned gray again. In these three last cases, participants did not score any points. After the feedback was given to the participants, all targets disappeared for 750 ms and then reappeared for the next trial. Only successful trials were kept for kinematics analysis of Experiments 1 and 2. This represents 85.3% ± 7.49 (mean ± SD) of trials for Experiment 1 and 81.5% ± 10.76 (mean ± SD) of trials for Experiment 2. However, the results obtained are similar when all trials are considered. For analyses concerning the adaptation to the force field, all trials exposed to the force field were kept, even if they were too slow or fast.
Main Experiment 1
This first experiment was an adaptation of an experiment performed by De Comite et al. (2021). In Experiment 1 (Fig. 1A), participants were instructed to perform reaching movements from the home target (2.5 cm × 2.5 cm) to the goal target located at 20 cm in the y-direction that could be either a small square (2.5 cm × 2.5 cm) or a wide rectangle (30 cm × 2.5 cm). The wide axis of the goal target was orthogonal to the straight-line path from the home target to the goal target.
During one trial, participants could face two types of online perturbations, either separated or combined: a lateral step load applied by the robot on their hand or a visual change in goal target shape, called target switch. The magnitude of the step load was ±9N (rise time of 5 ms), aligned with the x-axis. This force was triggered based on a hand position criterion, i.e., when the participant's hand crossed a virtual line parallel to the x-axis and located 6 cm from the starting point. It was switched off at the end of the movement (after the stabilization period). The target switch was triggered at the same time as the step load and consisted of an instantaneous change of the target shape, from a square to a rectangle, or vice versa.
The experiment was divided into eight blocks of 58 trials. Each block included 18 trials without target switch or step load (9 trials per target type: square or rectangle), 16 trials with only a step load [4 trials per combination of target type (2) and step load direction: left or right], 8 trials with only a target switch (4 trials per switching target: square to rectangle and rectangle to square) and 16 trials with both a step load and a target switch [4 trials per target switch type (2) and per step load direction (2)]. In total, participants performed 464 trials, including 32 trials for each combination of perturbations (except for the no-perturbation trials, where 72 trials were included). A score corresponding to the number of successful trials was projected on the virtual display to motivate participants. All these trials were randomly presented to the participants except for the first block, which began with eight trials toward a square or a rectangle without target switch or step load, so the participants could start with easy trials. Before these eight blocks, participants performed a training block composed of 24 trials, two of each type. In total, the experiment took 2 h including explanation and preparation of the participant for the experiment.
Main Experiment 2
The second experiment was very similar to the first one, with the addition of a velocity-dependent force field (FF; Fig. 1B). This velocity-dependent force field was applied to the hand of the participants in addition to the other perturbations. This force field consisted of a lateral force to the right proportional to the forward hand velocity:
In this experiment, participants performed eight blocks of 64 movements. The distribution of trials was the same as in Experiment 1, with six additional catch trials per block, three toward a square, and three toward a rectangle. Before the experiment, participants were exposed to a training block without FF, as for the first experiment. The first block started with eight trials toward a square or a rectangle without other perturbations than the velocity-dependent force field. This was done to avoid the first trials with step load and/or target switch being missed. The experiment also took 2 h in total.
Control experiments
Three control experiments were designed to decompose all the different factors present in Experiment 2. This first control experiment was a classical adaptation experiment in which participants had to perform reaching movements while exposed to a velocity-dependent force field similar to the one used in Experiment 2 (Fig. 1C1). Again, participants started from the home target (2.5 cm × 2.5 cm) and aimed for a similar goal target presented as a square at the same location and identical to the one used in Experiments 1 and 2 (2.5 cm × 2.5 cm, 20 cm in the y-direction). The second control experiment (Fig. 1C2) was the same as the previous one with the addition of target redundancy with or without a switch. Indeed, six participants experimented with a rectangle goal target in half of the trials and six participants experimented with the goal target which was a square switching to a rectangle in half of the trials (the other half being a square target). As the results in terms of motor adaptation for these two experiments were similar, the two experiments were merged. The last control experiment was also a classical adaptation experiment with the addition of a step load to the left or the right. These step loads were identical and in the same proportion as in Experiment 2 (Fig. 1C3).
In the three control experiments, participants started with a first block of 30 trials without force field to get used to the task. Then, they performed five blocks of 50 trials in the presence of the force field (and the corresponding additional perturbations). Each of these blocks included five catch trials, randomly inserted (10% of trials). Because we were interested in feedback corrections, we did not use error clamps or force channels, as these types of trials constrain movement and suppress natural corrective responses, limiting our ability to assess how participants responded to perturbations. At the end of the experiment, participants performed a wash-out block of 30 trials without force field. The experiment took 1 h in total.
With the main and the control experiments together, we can analyze, on the one hand, the influence of motor adaptation on policy update by comparing the results of the main experiments, and, on the other hand, the influence of target switch and step loads on motor adaptation by comparing Experiment 2 with the data from the different control experiments.
Data analysis
Kinematics data were recorded with a sampling rate of 1 kHz using Kinarm's Dexterit-E software (version 3.9). Preprocessing of the data was done using custom MATLAB scripts (MATLAB 2022a, MathWorks). Kinematics data were filtered with a dual-pass third-order low-pass Butterworth filter with a cutoff frequency of 20 Hz. The rest of the analyses were done using Python 3.11 via Anaconda Software Distribution (version 23.7.4; Anaconda 2023). All kinematic data were aligned with the onset of the step load (which also corresponds to a position threshold, as the step load is always applied when the hand crosses a virtual horizontal line located at 6 cm from the home target). We defined the final hand position as the position of the hand when the velocity dropped below a threshold set to
For the main experiments, the activity of two muscles of interest was measured. The two chosen muscles were the pectoralis major (PM) and the posterior deltoid (PD) in the right shoulder, as these are the main muscles recruited to respond to the left and right step loads in this configuration. Before placing the electrode over the muscle belly, the skin of the participants was cleaned with cotton wool and disinfected with medical alcohol. Gel was applied to the electrodes to improve signal conduction.
Muscular data were sampled at 1 kHz and amplified with a factor
When analyzing the data, we primarily focused on trials where participants experienced a step load to the right, in the same direction as the FF. This approach was taken because, when participants were exposed to a leftward step load, as the FF is directed to the right, the effect of the step load was attenuated, thereby reducing the statistical power of the analysis. However, these leftward perturbed trials played a critical role in preventing participants from anticipating the step load direction before movement. Moreover, for the kinematics, we selected the subsets of data that most directly addressed our central hypothesis and focused on trials with a target switch from a square to rectangle. This allowed evaluating the ability of participants to take advantage of the target redundancy.
Statistical analysis
The statistical difference between groups across Experiments 1 and 2 (with or without FF) for the final hand position and the difference in variance was computed using Student's t tests (implemented via the scipy.stats module from the SciPy library, v.1.11.1). The difference was considered significant when p < 0.05 as we made an intersubjects comparison and used EMG data.
To evaluate feedback responses depending on the change in target shape (square vs square to rectangle target) in both groups, we analyzed normalized EMG data. We only looked at the stretch response in the agonist muscle to the direction of the step load. The EMG data of each participant were averaged across trials of the same condition. Data were then averaged with nonoverlapping bins of 25 ms width. This bin width was the same as the one usually selected in other similar studies (Kurtzer et al., 2008; Nashed et al., 2012; De Comite et al., 2021).
A linear mixed model was used to assess the statistical difference in muscle activation and in final position. The general equation of this linear mixed model is as follows:
For the characterization of motor adaptation when the experiments involved a force field, we extracted the maximal deviation, which is the most eccentric position in the direction of the force field. This maximal deviation usually decreases with exposure to the FF and can be fitted with an exponential regression. A bootstrapping technique was used to resample participants (with replacement) one million times. The exponential regression model:
Results
Experiments 1 and 2: influence of the FF on flexible control
We first compared the two main experiments to assess whether exposure to the FF affected participants’ ability to react to online changes (visual or mechanical). In the first experiment, participants were exposed to changes in target shape during movement and step loads applied either leftward or rightward. The second experiment mirrored the first one but included a clockwise velocity-dependent FF. Figure 1, D and E, shows the force profile applied to the hand during movement in the case of a right or left step load for Experiments 1 and 2, respectively. Figure 2, A, B, F, and G, depicts the kinematics traces corresponding to rectangle and square goal targets without any target switch. As previously reported (De Comite et al., 2021), participants’ behavior in Experiment 1 depended on the shape of the target. A significant difference in intrasubject variance was observed between the square target (endpoint variance = 0.127) and the rectangle target (endpoint variance = 1.704, t(19) = 3.278, p = 0.004, paired t test). This larger variance for the rectangle target was indicative of the participant's tendency to exploit target redundancy.
Experimental paradigm. A, Illustration of targets and forces for the first main experiment. Participants were asked to perform reaching movements of 20 cm between the home target and the goal target, which could either be a rectangle or a square. For trials with target switch, the change occurred during the movement (from rectangle to square or vice versa). For trials with mechanical perturbations, rightward or leftward step loads were applied during the movement. The onset of the target switch and step loads is represented by the dotted line at 6 cm in the y-direction from the home target. This line was not visible to the participants, who saw only targets and a cursor representing the position of their hand. B, Schematic of the second main experiment. The instructions were the same as for the first experiment but in this case, participants were exposed to a velocity-dependent force field
Kinematics. A–C, Mean x–y traces across trials for each participant (light traces) and across participants (thick trace) with the FF (Exp. 2, purple) or without FF (Exp. 1, blue) from trials in which there was no step load for (A) a rectangle goal target, (B) a square goal target, and (C) a square goal target initially which became a rectangle during movement. D, Evolution of the x-position (mean across participants ± SE) over time during the movement with FF (purple) or without FF (blue) for a rectangle target (solid line), a square target (dotted line), or a switch from a square to a rectangle target (dashed line). Note that the curves for the square target and the switch from a square to a rectangle target are nearly superimposing each other. E, Evolution of the y-position (mean across participants ± SE) over time during the movement for all conditions. Note that all the curves are nearly superimposing each other F–J. Same as panels A–E for trials with a right step load.
Similar observations were made regarding the endpoint variance when participants performed movements in the presence of the FF, with the intrasubject variance remaining significantly different (endpoint variance for the square = 0.123, endpoint variance for the rectangle = 2.983, t(19) = 4.999, p = 7.97 × 10−5, paired t test). The difference in endpoint variance was even more pronounced in this case as the FF deviated the participants’ hand to the right. The x-position during movement increased more (Fig. 2D) and participants exploited the target's redundancy by letting their hand deviate along the target axis as indicated by the increase in variance in this direction. Together, these increases in final x-position and variance indicate a motor strategy that exploits the structure of the task. A more detailed analysis of the final position and the evolution of variance is presented later (Fig. 3).
Exploitation of target redundancy. A, B, Average final x-position along the rectangle width without force field (blue) and with force field (purple) following a right step load for each trial (mean across participants ± SE) when (A) the goal target was a rectangle and there was no visual switch and (B) the goal target switched from a square to a rectangle. C, D, Evolution of intrasubject variance with time (mean across participants ± SE) when (C) there was no visual switch of the rectangle target and (D) the goal target switched from a square to a rectangle. E, F, Mean of the difference between the maximal and the final variance for each experimental group when there is (E) no switch of the goal target and (F) a switch from square to rectangle. Positive values mean the endpoint variance was smaller than the peak variance across trials.
We can then look at the movements in which participants are exposed to rightward or leftward step load with or without FF. The endpoint variance was again higher for the rectangle target than for the square target for all conditions of step loads being with or without FF. For left step load without force field (endpoint variance square = 0.141, rectangle = 5.58; t = 6.826, p < 4.657 × 10−10, paired t test) and with force field (endpoint variance square = 0.104, rectangle = 4.74; t = 6.646, p < 2.34 × 10−6, paired t test). For right step load, without force field (endpoint variance square = 0.146, rectangle = 5.653; t(19) = 6.486, p < 4.23 × 10−6, paired t test) and with force field (endpoint variance square = 0.112, rectangle = 6.82; t(19) = 7.791, p = 2.48 × 10−7, paired t test). These findings collectively indicate that when the target was a rectangle, participants took advantage of target redundancy, resulting in higher endpoint variance.
No differences were observed in the y-position across conditions (Fig. 2E,J). However, regarding the x-position when participants were exposed to a rightward step load (Fig. 2I), the deviation in the positive x-direction was higher at the beginning of the movement with the FF. However, the final x-position was higher without FF, regardless of whether the target is a rectangle or a square transforming into a rectangle during movement. This is indicative of a bigger correction toward the center of the rectangle target in the presence of the FF.
We can then focus on the final position in the target. Under a step load without a target switch, participants effectively exploited target redundancy. Specifically, under rightward step loads, participants utilized target redundancy similarly, irrespective of the presence of the force field. This observation serves as a control condition: even if the net total force was different in Experiment 2 as the resultant was the sum of the force field and step load, our data indicated that participants exploited target redundancy in the same way when there was no target switch. Based on this condition, the effect of the switch could be interpreted in spite of different total forces across experiments. Figure 3A shows the evolution of the final position across all trials in this condition (rectangle target, no switch, right step load). A reduction in endpoint eccentricity across trials was observed. A linear mixed model was fitted to further characterize the evolution of the final position used as a dependent variable as a function of the trial number and group as fixed predictors and a random offset for each participant (see Materials and Methods). Only the final positions of valid trials were considered. For the trials toward a rectangle target without a switch, the effect of the trial number was significant (β1 = −0.05, CI 95% [−0.082, −0.018], z(1131) = −3.091, p = 0.002), with no significant effects observed for the group (β2 = 0.233, CI 95% [−0.593, 1.060], z(1131) = 0.554, p = 0.580) or the interaction between groups and trial number (β3 = −0.038, CI 95% [−0.083, 0.007], z(1131) = −1.664, p = 0.096). However, when a switch from a square to a rectangle target occurred (Fig. 3B), both the trial number (β1 = −0.055, CI 95% [−0.08, −0.03], z(1121) = −4.362, p = 1.29 × 10−5) and the group (β2 = −1.019, CI 95% [−1.667, −0.37], z(1121) = −3.079, p = 0.002) had significant effects, with no significant interaction (β3 = −0.022, CI 95% [−0.057, 0.014], z(1121) = −1.198, p = 0.231). Thus, the final position on the target was significantly less eccentric when participants encountered both the FF and a visual switch in the target shape. Interestingly, this difference in the final position on the target was not observed without a target switch, suggesting a reduced ability of participants to take advantage of target redundancy because of the combination of sudden change in target shape and exposition to the FF and not due to the FF alone.
A reduced ability to exploit target redundancy was also highlighted by the temporal profile of the variance along the x-axis. Indeed, a reduction in variance reflects the use of a state-dependent control policy that steers the hand back toward a target state (Todorov and Jordan, 2002b). Interestingly, such a decrease in intrasubject variance was observed when participants encountered the combination of FF and target switch (Fig. 3D) but not when there was no switch in target (Fig. 3C). We extracted the difference between the maximum variance across trials for each participant and the endpoint variance in this scenario. A positive value indicates that the variance peaks during movement and decreases near the endpoint, whereas a value of 0 indicates that the peak variance is observed at the endpoint. We observed that the endpoint variance was significantly reduced when participants were exposed to the FF in the switch condition (t(38) = 2.535, p = 0.0077, one-tailed independent t test; Fig. 3F), whereas no significant difference was observed without the target switch (t(38) = −0.118, p = 0.453, one-tailed independent t test; Fig. 3E). This further illustrates participants’ reduced ability to exploit target redundancy as the reduction in variance when exposed to FF and target switches indicated a tighter control of lateral coordinates in spite of the target redundancy.
Muscle activities related to step loads were analyzed by comparing mean activity during perturbed trials to that during unperturbed trials. Consistent with previous reports (De Comite et al., 2021), muscle activity toward a rectangle target was lower compared with a square target. In this study, the contrast of muscular activity between the movements toward a square or a square switching to a rectangle for Experiments 1 and 2 (Fig. 4A,D) was analyzed. In this analysis, the muscular activity in response to the FF has been removed when subtracting the muscular activity during trials without step loads. However, the raw muscular activity before step load is higher when participants are exposed to the FF (mean in a 50 ms window before onset of step load: 0.54 [a.u] without FF, 0.98 [a.u] with FF). This leads to gain scaling in short-latency stretch responses as the same muscle stretch will elicit larger responses for larger preperturbation muscle activity (Pruszynski et al., 2009). For Experiment 1, a clear difference in muscular activity emerged in the voluntary window (100–180 ms after force onset), which was less pronounced in the presence of the FF. The data were then grouped into 25 ms bins, aggregating the data over nonoverlapping 25 ms time windows. Using a linear mixed model on mean muscular activity across these 25 ms bins (Fig. 4B,E), a statistically significant difference was found from 100 to 200 ms (p < 0.005 for all bins). This difference was still present, although less significant, for some bins after 100 ms when the force field was present. By looking at the evolution of p values over a 25 ms sliding window, a noticeable drop could be seen after 100 ms without FF in Experiment 1 (Fig. 4C), whereas it was less clear with FF (Fig. 4F). The reduced statistical significance observed in the FF condition is due to a meaningful attenuation of the effect—consistent with our behavioral findings. Specifically, the movement kinematics showed that participants were still able to exploit target redundancy when they were exposed to the FF, only to a lesser extent. The EMG results mirrored this pattern. To avoid the arbitrary binning and consider correction for multiple comparisons, a permutation test was performed in addition. For both Experiment 1 and 2, within-subject difference was computed between the two conditions (square and switch from square to rectangle) at each time point. For each permutation iteration, the sign of the difference was randomly flipped for each participant, simulating the null hypothesis of no difference across subjects. Then, the mean difference across subjects was computed for each time point, building a null distribution of differences over 1,000 permutations. For each time point, a two-sided p value was computed by comparing the real mean difference to the null distribution, and the p value was corrected for multiple comparisons using the false discovery rate correction. The results of this analysis indicated a cluster of significant differences starting a bit before 150 ms without FF and no significant difference with FF. These results confirm our previous observation, showing a weaker effect of the observed difference when participants were exposed to the FF. Figure 4G illustrates distinct differences in muscular activity depending on the goal target during the long-latency and voluntary windows. In both experiments for the voluntary window, a gradual decrease in muscle activity was observed from the square target to the rectangle target. The muscular activity for a square target and a rectangle becoming a square was significantly higher than for a square becoming a rectangle or a rectangle, being without or with FF (even if it is less significant with FF). Interestingly, this clear modulation of the muscular activity depending on the target showed that participants were able to flexibly adjust their motor response. This indicates that there was no systematic upregulation of co-contraction used as a strategy to enhance limb impedance and robustness irrespective of the target geometry (Burdet et al., 2001; Crevecoeur et al., 2019). We further checked that participants exposed to the FF did not exhibit muscle co-contraction before step load onset. A significant difference was observed in pectoralis activity (t(38) = −2.559, p = 0.015, two-tailed independent t test) indicating a response to the velocity-dependent FF, already active before step load onset. However, no significant results were found for the deltoid activation (t(38) = −0.286, t = 0.777, two-tailed independent t test), suggesting no additional muscle contraction as the default strategy against perturbations.
Muscular activity. For the first main experiment (without FF), (A) muscular activity in the pectoralis major (solid line) and in the posterior deltoid (dotted line) following a right step load when the target was a square (dark blue) or a square which became a rectangle (light blue). B, Mean muscular activity over 25 ms bins for a square target (dark blue) or a target switching from a square to a rectangle (light blue). C, Log-scale evolution of the p value over a 25 ms sliding window. Time index corresponds to the timing of the end of the window. Red horizontal line represents p = 0.05. D–F, Same as A–C, but for the second main experiment, with FF. G, Mean muscular activity for each target condition during the long-latency and voluntary windows. H, Mean muscular activity across all trials in the pectoralis major and posterior deltoid in a window of 300 ms before the step load for the experiment without FF (blue) and with FF (purple).
These findings collectively indicate that participants were able to exploit target redundancy while adapting to the FF but that this ability was reduced when participants were exposed to the FF in addition to a target switch. The final position on the target was significantly less eccentric and a decrease in variance at the end of the movement was observed. These results were reproduced by looking at the muscular activity. The absence of systematic co-contraction suggests a modulation of control gains during movement at latencies that were a bit shorter than the ones previously reported (De Comite et al., 2021). These results therefore highlight a specific interference between flexible control (represented here by the target switch during movement) and motor adaptation.
Experiment 2 and control experiments: reduced extent of adaptation
Next, we sought to analyze the potential effect of changes in target redundancy on motor adaptation. To explore this, we compared the data from the second main experiment, where participants encountered online control and a velocity-dependent FF, to the data from the first control experiment featuring only the FF. The data from the second experiment were subsampled to extract only the trials toward a square target and analyze them to their actual trial indices. This allows comparing trials to the same target by taking into account the number of trials participants were exposed to the FF previously. The time axis was thus not influenced by this subsampling and so, a similar learning rate would indicate that the intermediate trials with other perturbations (of target or forces) did not slow down the adaptation. Figure 5A shows the learning curves from these experiments. Notably, the experiment incorporating both switches in target shape and step loads had twice the number of trials. It can be observed that participants reached a lower maximal deviation during standard FF adaptation. The after-effect quantified by the deviation in the direction opposite to the FF during catch trials was significantly more pronounced with the FF alone (t(30) = 3.868, p = 5.47 × 10−4, two-tailed independent t test; Fig. 5B). To quantify differences in learning rate and extent of adaptation, an analysis of the b and c parameters of the exponential regression (see Materials and Methods) was conducted (Fig. 5C). The bootstrap analysis revealed an overlap in the 95% HDI for the b parameter, indicating a similar learning rate across both experiments. However, there was no overlap in the 95% HDI for the c parameter indicating a reduced extent of adaptation in Experiment 2. The parameter a was not shown as it is directly linked to the c parameter. Indeed, the sum a + c corresponds to the starting point of the exponential curve, which was similar in both cases as evidenced by the overlap of 95% HDI for the distribution of a + c: [3.7, 4.94] in the FF condition and [3.35, 4.52] in the FF + step load + switch condition.
Motor adaptation. A, Maximal deviation (with respect to the straight line between the home and goal target) across trials for which participants are exposed to the force field for the experiment with target switch and step load in addition to the force field (purple) and the control experiment with the force field only (red). Trials toward a square target are represented against their index. B, Maximal deviation during catch trials represented against their index for both experiments (mean across participants + SEM; left part) and median deviation across the ten last catch trials (right part). C, Distribution of the b and c parameters of the exponential regression after 106 resampling of the population with replacement. Horizontal lines below the graphs represent the 95% HDIs. D, Comparison across all control experiments. Mean traces for the first (left) and last (right) 30 trials exposed to the force field for the experiment with the force field only (red), the experiment with target switch and step loads in addition to the force field (purple), the experiment with only target redundancy (rectangle and switch from square to rectangle, brown), and the experiment with step loads in addition to the force field (orange). Each line represents the mean for one trial across participants. E, Value of the c parameter of the exponential regression for the same four experiments. The 95% HDIs are projected on the y-axis.
Further experiments were warranted to elucidate whether this reduced adaptation stemmed from the target shape or the step load. To explore this further, two additional control experiments were conducted to separate the effects of the target shape and the step load. Figure 5D illustrates the mean trajectories for the first 30 and last 30 trials from Experiment 2 and from each control experiment. Initial trials exhibited substantial deviation, but varying levels of residual deviation toward the FF direction were observed in the 30 last trials across conditions. This was reflected by the c value of the exponential regression (Fig. 5E). The HDI overlap between the FF and FF + switch experiments indicated that the residual deviation and therefore, the extent of adaptation were similar in these two conditions. This indicated that changes in target shape did not influence the extent of adaptation to the FF. A dedicated analysis of behavior during learning with a rectangle target would require a follow-up study. However, an overlap of HDI for the FF + step load experiment with the corresponding interval in purple (Exp. 2) was observed. This overlap indicated that when we added a step load, the extent of adaptation was reduced.
Altogether, these results indicate that the learning rate is similar between the standard adaptation to the FF and the second experiment in which we added target switches and step loads. A difference in the extent of adaptation was however observed between these two experiments and further control experiments indicated that this difference in the extent of adaptation stemmed from the step loads.
We thus demonstrated an asymmetric interference between motor adaptation and flexible control (considered here as adjustments to online changes, i.e., switch in target or step loads). On the one hand, motor adaptation influenced flexible control by reducing participants’ ability to exploit target redundancy when the target switched from a square to a rectangle. On the other hand, flexible control (and more specifically step loads) reduced the extent of adaptation to the FF.
Discussion
This study investigated the interplay between flexible control and motor adaptation to determine whether these processes were completely independent or whether any interdependence between them could indicate possible overlap in their neural bases. We compared participants’ behavior in tasks involving either flexible control, motor adaptation, or both. Notably, we observed distinct behaviors when both mechanisms were engaged simultaneously compared with when they were engaged separately. We made the following key observations: first, when both mechanisms were combined, participants exhibited a reduced ability to exploit the redundancy of the rectangle target; second, they also demonstrated a lesser extent of adaptation to the force field.
The results from the first experiment were consistent with previous studies on the influence of target redundancy on reaching control policies (Nashed et al., 2012; De Comite et al., 2021). The difference in endpoint variance depending on the target shape indicated a modulation of the feedback gains depending on target redundancy. This modulation was further reflected in the EMG data, which showed a modulation consistent with the target structure and provided important measurements about the timing of this modulation. The latencies observed for EMG responses in Experiments 1 and 2, ∼100 ms, are comparable with the latencies of the visuomotor system (Franklin et al., 2017) and align with the findings that response amplitude to target jumps is sensitive to goal redundancy within ∼90 ms (Cross et al., 2019). Unexpectedly, these latencies were shorter than the ∼150 ms reported by De Comite et al. (2021), suggesting that the time needed to change the control policy could have been overestimated in this previous study potentially due to a smaller sample size. The analysis of participants’ ability to exploit target redundancy across trials revealed a progressive reduction in lateral endpoint eccentricity, regardless of the presence of a FF (Fig. 3A,B). This aligned with Orban de Xivry and Lefèvre (2016), who found that random target width changes reduced exploitation of target redundancy, possibly reflecting a switching cost. The effect was similar across conditions with or without FF, as reflected by the comparable slopes for lateral eccentricity evolution (Fig. 3A,B).
Regarding motor adaptation (Fig. 5, red curves), our results aligned with previous studies, showing an exponential decay of the maximal deviation in the FF direction (Shadmehr and Mussa-Ivaldi, 1994; Dizio and Lackner, 1995; Lackner and DiZio, 2005; Cluff and Scott, 2013). This observation was based on a statistical fit of a standard exponential model, in which we did not make any assumptions about the underlying dynamics. In contrast, other approaches were previously used in the literature. For example, state-space models (Smith et al., 2006; Kording et al., 2007) describe learning as a balance between error-dependent correction and baseline reversion, a sort of forgetting (Vaswani et al., 2015). Besides, the Kalman filter model integrates optimally sensory feedback and state estimates (Wei and Körding, 2010). While our approach focused on a statistical description, incorporating these dynamics-based models in future work could clarify the contributions of error correction, sensory integration, and baseline reversion, providing complementary perspectives on motor adaptation.
The main contribution of our work is to highlight a specific interference in two identified conditions. Although the experimental paradigm resembles a dual-task setup, the interference observed was not trivial and was specifically linked to motor control processes. Our first main observation was that the presence of a force field limited participants’ ability to modify their controller following target switches. This effect is unlikely to result from a general limitation in visual processing due to increased muscle activity, as previous work by Franklin et al. (2012) showed that background forces up to 6N did not modulate visuomotor responses. While the forces used in our experiment were somewhat higher, these findings suggest that increased muscle activation alone is unlikely to explain the observed interference. This observation provides indirect evidence that sensorimotor adaptation operates online as already suggested in previous studies (Crevecoeur et al., 2020a,b). The first main experiment showed that target redundancy exploitation was consistent across conditions with or without FF when the target did not change (Fig. 3A,C). Under the hypothesis that sensory prediction errors update an internal model of movement dynamics between trials (Shadmehr and Mussa-Ivaldi, 1994; Shadmehr et al., 2010), it is tempting to conclude that flexible feedback control and adaptation are unrelated. However, such an interpretation would miss the fact that interference became clearly apparent when the target switched during movement. This indicated that the interference mechanism occurred when the target switched, suggesting a specific timing for these perturbations to produce interference, that is when they both happened during an ongoing movement. The observations are consistent with our hypothesis that adaptation also engaged online processing (Kalidindi and Crevecoeur, 2023). The fact that interference occurred at a specific timing underscores the importance of measuring the timescales of these processes.
Our second key finding was the reduced extent of motor adaptation observed in Experiment 2. Our control experiment (Fig. 5D, orange curve) revealed that it was due to the superposition of different force perturbations (velocity-dependent FF and step load). Conceptually, these observations are consistent with the findings of Singh and Scott (2003) who showed that under uncertainty, participants relied on “local” sensorimotor associations, mapping loads sensed at one joint to control updates at the same joint. For instance, participants misattributed elbow loads as linked to elbow velocity, even when they were proportional to the shoulder. It was necessary to decorrelate the two signals with a different exposure protocol to enable proper association between shoulder velocity and elbow loads. Similarly, in our experiment, step loads added to the FF may have disrupted FF identification, suggesting interference about contextual parameters as generalized with the contextual inference model (COIN; Heald et al., 2021).
Our results can be interpreted within the framework of stochastic optimal control (SOC) initially introduced by Todorov and Jordan (2002b) to explain how humans achieve reliable success despite variations in motor behavior. In this model, changes in the target shape correspond to modifications of the cost function, while motor adaptation updates the state-space representation matrices. Mathematically, these two mechanisms therefore impact different sets of parameters (Kalidindi and Crevecoeur, 2023). However, our findings suggest interference between these mechanisms in practice. Specifically, when participants were exposed to the FF, they tended to apply a larger correction to the center of the rectangle target. This likely reflects an increase in feedback gains evoked when exposed to novel dynamical contexts (Franklin et al., 2017; Crevecoeur et al., 2019; Coltman and Gribble, 2020). Within the SOC framework, higher feedback gains imply a stronger correction in response to step loads. However, when a target switch was introduced in addition to the FF, the corrections became even larger. This suggests an interaction between the mechanisms governing the modulation of feedback gains and those involved in updating the controller following changes in the cost function. We could therefore hypothesize that a general cost function governs the update of these mechanisms, ensuring limited parameter adjustments for global efficiency. Alternatively, a hierarchical control structure may prioritize certain parameters over others, dynamically allocating resources based on task demands. Future modeling work should account for these shared resources and potential priority rules to better capture the interplay between those mechanisms.
Building on this understanding, our behavioral findings suggest that the neural structures underlying flexible control and motor adaptation may not be independent. While distinct neural circuits may underlie each process independently, the occurrence of interference when both are engaged suggests that the underlying neural processes are at least partially integrated or influenced by shared modulatory pathways. Although alternative explanations cannot be entirely ruled out, we propose several hypotheses regarding the possible neurophysiological basis of this behavioral effect. In particular, certain neural structures emerge as strong candidates based on the evidence and theoretical arguments presented above. One candidate locus of interference is the cerebellum, which is known to contribute to long-latency reflexes in response to unexpected mechanical perturbations (Strick, 1983; Kurtzer et al., 2013) and has been associated with sensorimotor adaptation (Shadmehr and Krakauer, 2008; Spampinato et al., 2017; Carey, 2024). Considering that afferent volleys of somatosensory feedback about the presence of both the force field and the step loads converge to the same cerebellar structure, signal superposition could disrupt how errors are mapped onto command updates.
Besides, the basal ganglia are often hypothesized to be involved in the representation of costs (Shadmehr and Krakauer, 2008), potentially influencing movement selection and change in controller depending on target shape. Bostan and Strick (2018) showed that basal ganglia and cerebellum form a densely interconnected network, with cerebellar motor output affecting basal ganglia input via the intralaminar thalamic nuclei. The reduced exploitation of target redundancy observed in our study during FF exposure may reflect a control policy shift driven by the basal ganglia, influenced by altered inputs from the cerebellum adapting to the FF. This interaction suggests potential interference originating from the interconnection between the basal ganglia and cerebellum, particularly if controller selection and motor adaptation are mediated by these two structures, respectively.
Another candidate locus of interference, suggested by the latency of ∼100 ms associated with the exploitation of target redundancy, is the pathway linking the visuomotor system to the sensorimotor network through the parietal cortex. Omrani et al. (2016) showed that parietal area A5 modulated control responses during task engagement. This region is also involved in visuomotor processing for limb motor actions (Kalaska, 1996). It could thus signal the change in target shape and mediate the subsequent change in control policy. Additionally, its role in proprioceptive limb position estimate (Rushworth et al., 1997) makes it a likely candidate for adaptation to a new environment in which the forces applied to the arm vary. However, further investigation is needed to understand the role of each of the structures involved in these motor control mechanisms and confirm the origin of the observed interference. Importantly, the timing of control updates, as well as the indication that interference occurred when these changes happened during movement, provide temporal constraints for future work on the computational and neural bases of these processes.
In all, this study introduced a novel paradigm for examining the interplay between task-dependent control and sensorimotor adaptation. This new approach allowed us to better understand how these two mechanisms interact and might help to develop more refined methods for studying their relationship in healthy and clinical populations.
Footnotes
The authors declare no competing financial interests.
A.D. was supported by Fonds de la Recherche Scientifique (F.R.S.-FNRS P.D.R. – Projet de Recherche). F.C. was supported by the F.R.S.-FNRS Grant 1.C.033.18. This work was additionally supported by a Concerted Research Action of Université catholique de Louvain (ARC; “coAction”) and by the WEL Research Institute (WELBIO Advanced Grant).
This is an open-access article distributed under the terms of the Creative Commons Attribution 4.0 International license, which permits unrestricted use, distribution and reproduction in any medium provided that the original work is properly attributed.
References
In this issue
