Abstract
Skilled motor performance depends critically on rapid corrective responses that act to preserve the goal of the movement in the face of perturbations. Although it is well established that the gain of corrective responses elicited while reaching toward objects adapts to different contexts, little is known about the adaptability of corrective responses supporting the manipulation of objects after they are grasped. Here, we investigated the adaptability of the corrective response elicited when an object being lifted is heavier than expected and fails to lift off when predicted. This response involves a monotonic increase in vertical load force triggered, within ∼90 ms, by the absence of expected sensory feedback signaling lift off and terminated when actual lift off occurs. Critically, because the actual weight of the object cannot be directly sensed at the moment the object fails to lift off, any adaptation of the corrective response would have to be based on memory from previous lifts. We show that when humans, including men and women, repeatedly lift an object that on occasional catch trials increases from a baseline weight to a fixed heavier weight, they scale the gain of the response (i.e., the rate of force increase) to the heavier weight within two to three catch trials. We also show that the gain of the response scales, on the first catch trial, with the baseline weight of the object. Thus, the gain of the lifting response can be adapted by both short- and long-term experience. Finally, we demonstrate that this adaptation preserves the efficacy of the response across contexts.
SIGNIFICANCE STATEMENT Here, we present the first investigation of the adaptability of the corrective lifting response elicited when an object is heavier than expected and fails to lift off when predicted. A striking feature of the response, which is driven by a sensory prediction error arising from the absence of expected sensory feedback, is that the magnitude of the error is unknown. That is, the motor system only receives a categorical error indicating that the object is heavier than expected but not its actual weight. Although the error magnitude is not known at the moment the response is elicited, we show that the response can be scaled to predictions of error magnitude based on both recent and long-term memories.
Introduction
Skilled motor performance depends critically on corrective responses that rapidly compensate for sensory prediction errors (i.e., mismatches between predicted and actual sensory feedback) that can arise from a variety of sources including noise in the motor commands and external perturbations. Corrective responses to perturbations during reaching movements have been extensively studied (Diedrichsen, 2007; Knill et al., 2011; Nashed et al., 2014; Diamond et al., 2015; Franklin et al., 2017. This work has shown that these responses, which can be observed in muscle activity changes within ∼60 ms, intelligently adapt to the physics of the arm and environment, the spatial and temporal accuracy requirements of the task, and decisions related to alternative corrective strategies (Scott et al., 2015; Scott, 2016). A focus of current models of motor control is on how these responses are adapted and optimized for different contexts (Scott, 2004; Todorov, 2004, 2009; Wolpert et al., 2011; Gallivan et al., 2018).
In many real-world tasks, the goal of reaching is to grasp and manipulate an object. Although corrective responses supporting object manipulation have been described (Flanagan et al., 2006; Johansson and Flanagan, 2009), little is known about the adaptability of these responses. When lifting an object, people predict its weight, and one of the most commonly evoked corrective responses in manipulation tasks occurs when the object is heavier than expected and fails to lift off at the expected time. In this scenario, the corrective response—an increase in lifting force that is initiated within ∼90 ms and continues until liftoff is sensed—is driven by the absence of expected sensory feedback (including tactile and other sensory signals) at the predicted time of liftoff (Johansson and Westling, 1988). This corrective response, which we refer to as the “corrective lifting response,” is fundamentally different from corrective responses to perturbations during reaching movements, which are driven by the presence of unexpected sensory feedback. Whereas these unexpected sensory signals provide information about the size and nature of the perturbation, such information is not available for responses triggered by the absence of expected sensory signals. When an object fails to lift off at the predicted time, the motor system receives a categorical error indicating that the weight is heavier than expected but not the actual weight of the object and hence the size of the perturbation (the difference between actual and predicted weight). Therefore, any adaptation of the corrective lifting response would have to be based purely on experience.
The aim of this study was to test the previously unexplored hypothesis that the gain of the corrective lifting response, as measured by the rate at which lifting force is increased, can be adapted based on expectations about weight variation developed from short- and long-term experience. In experiment 1, participants repeatedly lifted a test object (Fig. 1A,B) with a fixed baseline weight of 2 N. On occasional catch trials, the weight was increased to 4, 6, or 10 N in different blocks of trials. We hypothesized that participants would learn to predict the catch weight in each trial block and scale the gain of the corrective response to this occasional weight to preserve the efficacy of the response. In experiment 2, the baseline weight was fixed at either 3 or 7 N, and we examined a single catch trial with a heavier weight (20 N). On the assumption that participants learn, through long-term experience, that the variation in weight across similar objects increases with the mean object weight (e.g., kettles vary in weight more than cups), we hypothesized that the gain of the corrective response would scale with the baseline weight, again preserving the efficacy of the response. Confirmation of these predictions would uncover an important sensorimotor control mechanism that supports the class of motor tasks in which the sensory information available when a perturbation occurs does not specify the size and nature of the perturbation.
Apparatus and data scoring. A, Left, In experiments 1 and 2, participants grasped and lifted a test object with the tips of the index finger and thumb, which contacted sensors that measured the forces applied by the digits (left). Right, In experiment 3, participants held an additional 400 g tungsten disk in the hand while lifting the test object. B, The device used to control the weight of the object. A linear motor was used to position a trolley along a rotating rod, attached via a string running through pulleys, to a hook located in the center of the object. C, Vertical force (VF) and the first (VF Rate) and second derivatives of VF, and vertical position, velocity, and acceleration in a baseline trial in which the object weighed the expected 2 N (green) and a catch trial in which the object weighed 6 N (red). For the catch trial, the three vertical dashed lines mark the onset and offset (Lift-off) of the corrective response (gray zone) and the onset of the hold phase. D, Illustration of the method used to determine the amplitude and period of oscillations in VF during the corrective response.
Materials and Methods
Participants
In experiment 1, a group of 11 participants (6 females) completed both the 2-4 condition and the 2-6 condition, and a separate group of 8 participants (5 females) completed the 2-10 condition. In experiment 2, 10 participants (6 females) completed the 3-20 condition, and 12 participants (4 female) completed the 7-20 condition. In experiment 3, added as a control for experiment 2, 9 participants (6 females) completed the control condition (described below). Participants provided written informed consent and were compensated for their time. All participants were undergraduate or graduate students at Queen's University and had normal or corrected-to-normal vision. In experiments 1 and 2, participants used their dominant hand to lift the test object, and a modified version of the Edinburgh Handedness Inventory (Oldfield, 1971) was used to confirm their self-reported dominant hand. In experiment 3, participants alternately lifted the test object with their right and left hands, and the corrective lifting response was tested using the right hand, which was the dominant hand for all participants in this condition. A Queen's University ethics committee approved the study protocol.
Although we are unaware of any previous work examining how the gain of the corrective lifting response changes across contexts, many studies have examined corrective responses during goal-directed reaching movement made under different conditions. For example, a number of studies have investigated the gain of the corrective response made when the viewed position of the hand is displaced while reaching to various targets (Saunders and Knill, 2004, 2005; Knill et al., 2011; Dimitriou et al., 2013; Diamond et al., 2015; Gallivan et al., 2016; Franklin et al., 2017). In these studies, there are between 8 and 12 participants per experimental group. These sample sizes provide sufficient power to detect the large effects of interest, which can typically be observed in most, if not all, participants. In a previous study (Gallivan et al., 2016), in which we examined changes in the gain of corrective responses when reaching to targets of different width, we found an effect size of Cohen's d = 2.57. A power analysis using this effect size and a power of 0.95 indicated that five participants would be required (G*Power, version 3.1; Faul et al., 2009). To err on the side of caution, and to be consistent with previous work, we aimed to obtain at least eight participants in each experimental group. As shown below, all the key results presented here were very consistent across participants and clearly observed at the level of individual participants.
Apparatus
Participants sat in front of a tabletop and used their preferred hand to lift a test object located on the table surface (Fig. 1A). The surface was covered with a hard, rubberized mat with high friction, which prevented the object from sliding if small horizontal forces were applied. The test object consisted of a shell (hollow 5 cm3 cube with an open bottom) composed of the opaque black polyoxymethylene plastic Delrin. A handle, attached to the top of the object, was instrumented with force-torque (F/T) sensors (Nano17 F/T, ATI Industrial Automation) that enabled us to measure the forces applied by the thumb and index finger when lifting. The test object was also equipped with a position sensor (Polhemus LIBERTY).
To set the weight of the test object, we used a linear motor to position a trolley (0.8 kg) along a rotating carbon fiber rod (1 m long) that was attached via a string running through pulleys to a hook located in the center of the test object (Fig. 1B). By locating the trolley at different positions along the rod, we could specify the weight of the object with a resolution of 0.1 N. The linear motor was moved on every trial (i.e., including trials in which the weight was not changed) so that participants would not be provided with cues about object weight.
Procedure
At the start of each trial, a tone (400 Hz, 500 ms) cued the participant to reach toward, grasp, and then lift the object vertically to a height of 2.5 cm above the tabletop. Participants were instructed to lift confidently and quickly after establishing contact between the fingertips and the object. A second tone (200 Hz, 500 ms), delivered 1000 ms after the object lifted off, signaled the participant to replace the object on the tabletop. In between trials the hand was located ∼5 cm in front of the object. At the end of each lift trial, visual feedback was provided about the height of the lift for 1500 ms. In particular, we displayed both the target height and the maximum lift height as horizontal lines on a display located behind the object. Each lift trial lasted ∼5000 ms, and the interval between successive lifts was ∼2000 ms.
In experiment 1, all participants first completed 20 practice trials (i.e., lifts) with the 2 N baseline weight. One group of participants then completed two blocks of 110 lifts, one block in the 2-4 condition and one in the 2-6 condition with the order counterbalanced across participants. Participants took a short (1–2 min) rest break between blocks. A second group of participants completed a single block of 110 trials in the 2-10 condition. In all three conditions, the object weight was set to 2 N for the first 10 lifts. In the following 100 lifts, the weight was set to 2 N in 80% of trials, and in the remaining 20% of trials, the weight was increased to either 4 N (2-4 condition), 6 N (2-6 condition), or 10 N (2-10 condition). Specifically, the weight was increased in 2 trials randomly selected from each successive block of 10 trials subject to the constraint that catch trials were separated by at least 2 baseline trials. Note that, following the practice trials, participants were told that the weight of the object could change on occasional trials.
In experiment 2, two groups of participants first completed 20 consecutive baseline lifts with the weight set to 3 N (3-20 condition) or 7 N (7-20 condition). For both groups, the weight was then unexpectedly increased to 20 N for a single catch trial. Participants were not informed about, and could not anticipate, the catch trial.
In experiment 3, one group of participants completed a single condition as a control for experiment 2. The aim of this condition, which we refer to simply as the Control condition, was to test for peripheral gain scaling in the corrective lifting response. More specifically, we tested whether there is an obligatory increase in the gain when additional torques, unrelated to the test object, must be generated at the joints involved in lifting. In the Control condition, participants first performed 26 consecutive baseline lifts with the weight of the test object set to 3.2 N. These 26 lifts consisted of alternating lifts using the right and left hands starting with the right hand, which was the dominant hand for all participants in experiment 3. On the next right-hand lift (trial 27), the weight was unexpectedly increased to 14 N for a single catch trial. Critically, throughout the experiment, participants held a 400 g (3.92 N) tungsten disk in their right hand. The diameter and height of the disk were 6.3 and 0.7 cm, respectively. The disk was grasped with the metacarpophalangeal joints of the index finger and thumb so that the test object could still be lifted with the tips of the index finger and thumb (Fig. 1A). We had participants perform alternating right- and left-hand lifts so that they received strong contextual information distinguishing the unchanging weight of the test object from the weight of the disk held in the right hand. Note that in baseline lifts with the right hand, the total weight being supported at the time of object liftoff was 7.12 N, which slightly exceeds the total weight in the 7 N condition. As a consequence, the joint torques—including the torques at the shoulder, elbow, wrist, and metacarpophalangeal joints—required to lift the 3.2 N object in experiment 3 were similar to those required to lift the 7 N object in experiment 2. After each baseline lift in experiment 3, participants viewed a time-varying trace of their vertical force (VF), starting 100 ms before the time vertical force started to increase and ending 850 ms later, on a display. The display included horizontal lines at 0, 2.7, and 3.7 N and a highlighted box located between 200 and 300 ms after the onset of vertical force increase (on the x-axis) and between 2.7 and 3.7 N (on the y-axis). With this feedback, we were able to obtain lifts in the Control condition that were similar, in duration and height, to those obtained in experiment 2.
Data processing and analysis
Data from the force-torque sensors and the position-angle sensor were sampled at 250 Hz and digitally smoothed using a fourth-order, zero phase lag, low-pass Butterworth filter with a cutoff frequency of 14 Hz. We defined the VF as the sum of the VFs applied to the two sensors. The first and second derivatives of VF with respect to time and the vertical velocity and acceleration of the object were computed using a first-order central difference equation. Figure 1C shows the time-varying kinematic and force signals for a baseline trial with a 2 N weight and a catch trial with a 6 N weight.
The onset of the predictive phase of the lift was deemed to begin when the VF rate last exceeded 0.1 N/s before reaching its maximum value. In baseline trials, the end of the predictive phase was defined as the moment of liftoff. In catch trials, the end of the predictive phase was defined as the time at which the VF reached the expected baseline weight of the object. The expected baseline weight was computed separately for each participant in each condition as the mean VF at liftoff in all the baseline trials. (Note that we computed the expected baseline weight for each participant and condition to compensate for any slight changes in weight because of calibration.) When lifting objects just off a surface, people produce a VF rate profile—during the predictive phase—that is roughly bell shaped (Fig. 1C). To quantify the rate at which VF increased during the predictive phase of the lift, we used the VF rate observed at the time the VF reached half the baseline weight rate of the object (e.g., 1 N in experiment 1). The vertical dashed cyan line in Figure 1C indicates the time at which the VF reached 1 N in both trials shown. For both trials, the VF rate at this time was close to the peak VF rate. (Note that we used the VF rate at the time VF reached half the object weight rather than the peak VF rate because the VF rate profile can occasionally have two peaks.) Because people typically scale the rate at which they increase VF to the expected weight of the object (Johansson and Westling, 1984; Flanagan and Beltzner, 2000; Flanagan et al., 2008), the VF rate at the time VF reached half the weight of the object provides an index of expected weight. For simplicity, we refer to this measure as the “predictive VF rate.”
To determine the onset of the corrective response, in each catch trial we identified the time at which the VF clearly started to increase after the initial, predictive increase (targeted for the baseline weight). The time of this point was determined by finding the local maximum in the second derivative of VF with respect to time (Fig. 1C, cyan circle). The time of liftoff was taken as the end of the corrective response. The gray zone in Figure 1C shows the corrective response. To quantify the gain of the corrective response, we computed the corrective VF slope, defined as the increase in VF, during the response, divided by the duration of the response. To estimate the latency of the corrective response, for each catch trial we determined the time at which the VF reached the expected baseline weight (computed separately for each participant and condition as described above). This expected liftoff time was then subtracted from the onset of the corrective response.
Note there are two reasons we could not use the standard statistical approach to compute response latency based on determining when signals from trials with and without responses (or trials with responses in different directions) significantly differ in time. First, at the moment an object lifts off, there can be a decrease in VF rate that is related to the transition from isometric to shortening contractions of the lifting muscles (Johansson and Westling, 1988). As a consequence, the VFs in perturbed trials and unperturbed trials can differ at the moment of expected liftoff. Second, when lifting an object, people increase VF to a target level that is a little greater than the expected weight so that the object will accelerate upward off the surface. Even when repeatedly lifting an object with the same (and therefore predictable) weight, both this target VF and the rate of change of VF can vary considerably. This variability is problematic for the standard statistical approach that relies on there being very little variability before the perturbation.
For each catch trial, we also determined the onset and offset times of the hold phase when the object was held aloft—at a more or less constant height—following liftoff. When the object is lifted, the vertical position increases and, in almost all trials, slightly decreases before being held steady. This results in an initial positive peak in vertical velocity followed by a negative peak. The onset of the hold phase was taken at the time at which the velocity increased above −1 cm/s following the negative peak. When the object is replaced on the surface, there is a large negative peak in vertical velocity. The offset of the hold phase was taken as the time at which the velocity last decreased below −1 cm/s before the negative peak. The onset of the hold phase in the 6 N catch trial shown in Figure 1C is indicated by a vertical dashed line. (The offset occurred ∼600 ms later and is not shown.)
As noted in the original study describing the corrective lifting response (Johansson and Westling, 1988), and as can be observed in Figure 1C, the corrective increase in VF involves an oscillatory component. To quantify the amplitude and frequency of this component, we used the following method. We first high-pass filtered the VF rate with a cutoff frequency of 4 Hz to remove any low-frequency components, including the baseline offset corresponding to the overall increase in VF. The circles in Figure 1D (top) shows the filtered VF rate for the corrective response in the 6 N trial shown in Figure 1C. The filtered VF rate was then partitioned into a series of segments (alternating orange and cyan circles) defined by successive peaks and inflection points identified by zero crossings in the first and second derivatives of the VF rate, respectively (Fig. 1D, gray vertical lines). For each segment, we fit a quarter of a sinusoid matching the amplitude and duration of the segment (Fig. 1D, alternating orange and cyan curves). The thin black curves in Figure 1D (middle) show the separate sinusoids for the seven segments in this particular trial. The thicker purple sinusoid represents the estimated filtered VF rate based on the mean amplitude and mean frequency of the segment sinusoids. Figure 1D (bottom) shows the time integral of the estimated filtered VF rate and represents the estimated oscillation in VF. The amplitude (solid vertical line) and period (dashed horizontal line) of this estimate were computed for each trial. Finally, for each catch trial, we also used this method to estimate the amplitude and period of oscillations in VF during the hold phase.
Note that we also examined grip force (GF) applied normal to the contact surfaces in the horizontal plane. Consistent with previous work (Johansson and Westling, 1988; Danion and Sarlegna, 2007; Diamond et al., 2015), we found that GF was modulated in synchrony with VF during both the initial predictive increase in VF and, in catch trials, the corrective increase in VF. However, we focused our analysis on VF because it is this force that is both predicted and controlled by the sensorimotor system to accomplish the primary goal of the task (i.e., lifting the object). Moreover, although measures based on GF showed the same patterns as measures based on VF, they are far more variable because GF depends not only on VF but also on the friction between the contact surface and the digits and grip force safety margins, both of which can vary substantially across participants.
Experimental design and statistical analysis
In experiment 1, one group of participants completed the 2-4 and 2-6 conditions, and we used pair tests and repeated-measures ANOVAs to compare these two conditions. Another group of participants completed the 2-10 condition, and we used independent samples t tests to compare this this condition with the 2-6 condition. In experiment 2, separated groups of participants completed the 3-20 and 7-20 conditions, and we used independent samples t tests to compare these conditions. In experiment 3, a single group of participants completed the Control condition, and we used independent samples t tests to compare this Control condition with the 3-20 and 7-20 conditions from experiment 2. A p value < 0.05 was considered to be statistically significant.
Results
Experiment 1
In the first experiment, participants repeatedly lifted the test object that weighed 2 N in the majority of the trials, which we refer to as baseline trials. In occasional catch trials, the weight of the test object was increased from the baseline weight of 2 N to a heavier weight. Within an experiment condition, the heavier weight was kept constant and was set to either 4, 6, or 10 N. One group of participants completed the 2-4 and 2-6 conditions in which the heavier weight was 4 or 6 N. A second group of participants completed the 2-10 condition in which the heavier weight was 10 N.
Figure 2 shows VF, VF rate, and vertical position traces for all lifts from a single participant who completed the 2-4 and 2-6 conditions and from a different participant who completed the 2-10 condition. For illustrative purposes, for each participant we have highlighted a single, exemplar lift for each weight (thicker lines), as well as the first catch trial for each heavier weight (dashed lines). As expected, the amplitude and rate of the initial predictive increase in VF was similar in baseline and catch trials. In both baseline and catch trials, the increase in VF targeted the 2 N baseline weight. On all catch trials, the failure of the object to lift off at the expected time triggered a corrective response, that is, an increase in VF, that ultimately led to liftoff. As illustrated in the figure, the gain of the response quickly scaled after the first one or few catch trials to the catch trial weight. That is, the slope of the corrective increase in VF tended to be greater in the 2-6 condition than in the 2-4 condition and greater still in the 2-10 condition. This scaling is also evident in the VF rate traces, where the average rate during the corrective response increases with the catch weight. The fluctuations in the VF rate traces clearly illustrate the oscillatory component of the corrective response. Note that the VF rate traces also illustrate oscillations in VF during the hold phase of both baseline and catch trials.
Representative trials. A–C, VF, VF rate, and position traces from all trials performed by a participant who completed the 2-4 and 2-6 conditions. For illustrative purposes, we highlighted a single representative trial for each weight (thicker lines) and the first catch trial in each condition (dashed lines). D–F, Corresponding data for a participant who completed the 2-10 condition. Note that the dashed yellow line in F does not increase above zero because in the first lift the object did not lift off within 800 ms of the start of the corrective response. For each participant, the catch trials (blue, red, and yellow traces) are aligned to the onset of the corrective response (A–F). The baseline trials (green traces) were aligned so that the time at which VF reached 1 N matched the mean time at which the force reached 1 N in catch trials.
For all trials, we determined the predictive VF rate—the VF rate at the time VF reached half the baseline weight of the object—as an index of expected weight. To assess whether participants updated their expectation of object weight following a catch trial, and whether such updating would change during the experiment, we examined the increase in predictive VF rate from the catch trial to the next baseline trial for each of the first 20 catch trials. Note that for this analysis, we combined all participants and only included the first condition (i.e., the first 20 catch trials) experienced by participants who completed the 2-4 and 2-6 conditions. Figure 3A (top) shows the change in predictive VF rate from each catch trial to the first baseline trial following the catch trial (first Baseline trial after Catch – Catch). A small but significant increase in predictive VF rate was observed for the first catch trial (t(18) = 3.77, p = 0.001), but not in any subsequent catch trials. Moreover, the updating on the first catch trial was momentary. Figure 3A (middle) shows the change in predictive VF rate from each catch trial to the second baseline trial following the catch trial (second Baseline trial after Catch – Catch). No significant differences in predictive VF rate were observed. Finally, as shown at the bottom in Figure 3A, and as expected, for all catch trials, there was no significant change in predictive VF rate from the baseline trial before the catch trial to the catch trial (Catch – Baseline trial before Catch).
Experiment 1 results. A, Differences in predictive VF rate between catch trials and subsequent and prior baseline trials as a function of catch trial number. B, Corrective VF slope as a function of catch trial number in each condition. The dots show group means, and the height of the shaded area indicates ±1 SEM. The blue/red and red/yellow stars indicate significant differences (p < 0.05) between the 2-4 and 2-6 conditions and the 2-6 and 2-10 conditions, respectively. C, Corrective VF slope on the first and last catch trials in the 2-4 and 2-6 conditions shown for each condition order. D, E, Corrective VF slope (D) and total corrective response duration (E) over the last 10 catch trials in each condition. F, Lift height for each object weight over the last 10 trials in each condition. Bars show group means, error bars indicate SEM (D–F), and lines and dots show individual participant means (D, E) or medians (F). The symbols *, **, and *** indicate significant levels of p < 0.05, p < 0.01, and p < 0.001, respectively. The letters n.s. indicate non significance (p > 0.05).
For all catch trials, we determined the latency of the corrective response, that is, the time between expected liftoff and the onset of the response. For each participant, we computed the mean latency across all catch trials (40 trials for participants who completed the 2-4 and 2-6 conditions and 20 trials for participants who completed the 2-10 condition). There was no significant difference between the two groups of participants (t(17) = 0.105, p = 0.92), and across all participants the latency was 77.35 ± 3.22 ms (mean ± SE). This latency is roughly consistent with the latencies observed for other corrective responses in both reaching and object manipulation tasks (Johansson and Flanagan, 2009; Pruszynski et al., 2016; Scott, 2016).
To quantify the gain of the corrective response, for each catch trial we determined the slope of the increase in VF during the corrective response. Figure 3B shows the mean corrective VF slope, averaged across participants, as a function of the catch trial number for each condition. The corrective VF slopes across conditions were initially similar, they differed and scaled with the catch weight within a few catch trials. To compare the corrective VF slopes at each catch trial number, we conducted paired t tests to compare the 2-4 and 2-6 conditions, and independent samples t tests to compare the 2-6 and 2-10 conditions. The blue/red and red/yellow stars in Figure 3B indicate significant differences (p < 0.05) between the 2-4 and 2-6 conditions and the 2-6 and 2-10 conditions, respectively. Whereas the corrective VF slope increased in the 2-6 and 2-10 conditions, it remained roughly constant in the 2-4 condition, indicating that the default corrective VF slope, that is, the initial slope seen in all three conditions, was appropriate for the 2-4 condition.
Figure 3C shows the mean corrective VF slope for the first and last catch trials in the 2-4 and 2-6 conditions for participants who completed the 2-4 condition first (left) and participants who completed the 2-6 condition first (right). As illustrated in the figure, the order in which the conditions were experienced did not affect the initial and final corrective VF slope in either condition. Note that the corrective VF slope, that is, the gain of the response, appeared to reset to the default value when participants experienced the 2-4 condition, following a break, after the 2-6 condition. In contrast, in the 2-6 and 2-10 conditions in which the corrective VF slope was elevated above the default value, the gain was maintained during the experimental session.
To examine the effects of condition (i.e., catch trial weight) on the corrective response following adaptation (i.e., the steady-state response), we examined response parameters over the last 10 catch trials. Figure 3D shows for each condition the group mean corrective VF slope (based on participant means) over the last 10 catch trials. To evaluate the effects of catch trial weight on the corrective slope, we used paired t tests to compare the 2-4 and 2-6 conditions and used independent samples t tests to compare the 2-6 and 2-10 conditions. These planned comparisons revealed significant differences in corrective VF slope between the 2-4 and 2-6 conditions (t(10) = 11.90, p = 3.2e-7) and between the 2-6 and 2-10 conditions (t(17) = 3.95 p = 0.001). Figure 3E shows for each condition the average total duration (based on participant means) from predicted liftoff to actual liftoff of the corrective response over the last 10 catch trials. The total duration of the response increased from the 2-4 condition to the 2-6 condition (t(10) = 10.02, p = 2.0e-6) and from the 2-6 to the 2-10 condition (t(17) = 3.31, p = 0.004).
Figure 3F shows the lift height observed in baseline and catch trials in each condition over the last 10 catch trials (mean of participant medians). Note that we used participant medians to guard against occasional outliers. We found no difference in lift height between baseline and catch trials in the 2-4 (t(10) = 1.73, p = 0.11) and 2-6 (t(10) = 2.06, p = 0.07) conditions, whereas the lift height was slightly smaller in catch trials compared with baseline trials in the 2-10 condition (t(8) = 4.34, p = 0.003). Importantly, in all cases the maximum lift height was close to the goal provided to participants (i.e., lift to a height of 2.5 cm). As can be seen in Figure 3F, the lift height in catch trials was quite similar across conditions. At first glance, we might have expected the lift height to increase with the catch trial weight because VF is increasing more rapidly at the time liftoff occurs, resulting in greater overshoot in VF before the corrective response is terminated. However, because the catch trial weight increases, the overshoot in VF would not necessarily lead to an overshoot in lift height.
As originally noted by Johansson and Westling (1988), and as can be seen in Figure 2, the overall increase in VF during the corrective response involves an oscillatory component. Oscillations in VF can also be observed during the hold phase. To characterize these oscillations, we computed, for each participant the average amplitude and period of the estimated oscillation in VF (Fig. 1D, method), during both the corrective response and the subsequent hold phase, over the last 10 catch trials. Figure 4A shows the estimated amplitude of VF oscillations as a function of condition and trial phase, that is, corrective response versus hold phase. For participants who completed the 2-4 and 2-6 conditions, a repeated-measures ANOVA revealed effects of condition (F(1,10) = 24.54, p = 6.0e-4) and trial phase (F(1,10) = 56.96, p = 2.0e-5) on amplitude, as well as a significant interaction (F(1,10) = 9.57, p = 0.01). For participants who completed the 2-10 condition, a paired t test demonstrated an effect of trial phase on amplitude (t(7) = 4.16, p = 0.004). The increase in amplitude with the catch weight during the hold phase is consistent with previous work on isometric force generation (Novak and Newell, 2017). We are not sure why the amplitude did not consistently increase with the catch weight (and hence the average VF) during the corrective response. However, in all conditions, the amplitude is substantially (two to three times) greater during the corrective response, despite the average VF being smaller than in the subsequent hold phase.
Characteristics of VF oscillations. A, Estimated amplitude of oscillatory component of VF across conditions and phases; that is, corrective response and hold phase. B, Estimated amplitude of oscillatory component of VF across conditions and phases. Thin lines and circles represent individual participants, and error bars indicate ±1 SEM.
Figure 4B shows the estimated period of VF oscillations as a function of condition and trial phase. For participants who completed the 2-4 and 2-6 conditions, there was an effect of condition (F(1,10) = 14.33, p = 0.004) but no effect of trial phase (F(1,10) = 4.26, p = 0.07) and no interaction (F(1,10) = 0.11, p = 0.74). For participants who completed the 2-10 condition, a significant effect of trial phase was observed (t(7) = 3.33, p = 0.01). The origin of these small but significant effects on the period is not clear to us. However, we would emphasize that, overall, the period is quite consistent across conditions and phases (ranging from 82 to 91 ms on average) and, in all cases, is within the range of physiological tremor. Such oscillations have been shown to be of central origin and are not driven by mechanical or peripheral feedback loop resonances (Vallbo and Wessberg, 1993; Wessberg and Vallbo, 1995, 1996; McAuley et al., 1997; Wessberg and Kakuda, 1999). Importantly, the corrective lifting response does not involve a series of distinct increases in VF, each triggered by the absence of sensory events signaling liftoff, in which case we would expect a delay between each increase; that is, the response does not involve intermittent control (Meyer et al., 1988; Harris, 1995). Rather, the corrective response involves a steady (albeit oscillatory) increase in VF that continues until liftoff is detected.
Experiments 2 and 3
In experiment 2, we tested the prediction that the gain of the corrective lifting response scales with the predicted weight of the object, that is, the expected weight in baseline trials. This prediction is based on the assumption that through experience people learn that the magnitude of errors in weight prediction tends to increase with predicted weight, that is, that the variation in weight of objects that are on average heavy (e.g., suitcases) is greater than the variation in weight of objects that are on average lighter (e.g., briefcases). Two groups of participants first completed 20 lifts of the test object with the baseline weight set to either 3 N (3-20 condition) or 7 N (7-20 condition) and then completed a single catch trial in which the weight was increased to 20 N. Figure 5A shows the VF trace from the catch trial for every participant in the 3-20 (orange traces) and 7-20 (purple traces) conditions. Because the catch weight was so large, the object often did not lift off following the initial corrective response, which for many of the participants continued for ∼400 ms. We observed a variety of behaviors after the initial response with some participants generating a second volitional lift (sometimes followed by a second corrective response) and others giving up before liftoff occurred. As illustrated in the Figure 5A, the slope of the initial corrective response was greater in the 7-20 condition than in the 3-20 condition.
Results from Experiments 2 and 3. A, B, VF traces for the single catch trial in the 3-20 and 7-20 conditions (A) and the Control (Ctr) condition (B). Each trace represents a single participant, and traces from all participants are shown. Traces are aligned to the start of the corrective response. C, Corrective VF slope over the first 200 ms of the corrective response (A, B, gray regions) in each condition. D, Predictive VF rate in each condition. C, D, Bars show group means, error bars indicate ±1 SEM, and dots show individual participants. The symbols ** and *** indicate significant levels of p < 0.01 and p < 0.001, respectively. The letters n.s. indicate non significance (p > 0.05).
To quantify the corrective response, for each participant we took the VF slope over the first 200 ms of the response (Fig. 5A, gray region) as the corrective response lasted for at least 200 ms in all participants. Consistent with our prediction, the corrective VF slope was greater (t(20) = 3.84, p = 0.001) in the 7-20 condition than in the 3-20 condition (Fig. 5C).
We also characterized the initial rate of VF increase during the predictive phase of the lift by computing for each participant the average VF rate at the time the VF reached half the baseline object weight across the first 20 trials. As expected, the VF rate at half the weight was much greater (t(20) = 4.98, p = 7.2e-5) in the 7-20 condition than in the 3-20 condition (Fig. 5D) as participants in both groups attempted to lift the object in around the same amount of time (Johansson and Westling, 1988; Gordon et al., 1991; Flanagan and Beltzner, 2000; Cole, 2008; Hermsdörfer et al., 2011).
We also estimated the amplitude and period of the oscillations in VF during the first 200 ms of the corrective lifting response in the catch trial. Independent samples t tests revealed a significant effect of condition on the amplitude of the VF oscillation (t(20) = 3.06, p = 0.006) but no significant effect on the period (t(20) = 0.34, p = 0.73).
We found no difference (t(20) = 0.44, p = 0.66) between the 3-20 and 7-20 conditions in terms of response latency; across all participants in both groups the latency was 95.1 ± 5.1 ms. Note that this latency is longer than the latency observed in experiment 1 (77.4 ± 3.2 ms). This result prompted us to examine the variability across catch trials in response latencies in experiment 1. For this analysis, we combined all participants, taking the 20 catch trials from the first condition experienced by participants who completed the 2-4 and 2-6 conditions. For each catch trial, we determined the mean response latency, averaged across all 19 participants. The average response latency across the 20 catch trials was 82.4 ms, and the SD was 7.3 ms. Thus, the response latency on the single catch trial in experiment 2 was within ±1.74 SDs of the response latencies observed across catch trials in experiment 1. (Note there was no correlation between response latency and catch trial number in experiment 1; r = 0.17; p = 0.48).
We would not expect the larger gain of the lifting response in the 7-20 condition compared with the 3-20 condition to be because of peripheral gain scaling associated with the greater forces and joint torques required to lift the 7 N object compared with the 3 N object. Work examining responses to arm perturbations has shown that gain scaling is only observed for short-latency (20–45 ms) spinal reflex responses, which are largely insensitive to context; gain scaling is not observed for long-latency (50–105 ms) supraspinal responses, which can be readily modified by context (Pruszynski et al., 2009). Thus, we would not expect obligatory gain scaling in the corrective lifting response, which is clearly a long-latency and context sensitive response.
Nevertheless, to examine the issue of peripheral gain scaling, in experiment 3 we ran an additional condition, which we refer to as the Control condition, in which we dissociated the torques required to lift from the weight of the object. Participants first completed 26 alternating right- and left-hand lifts (13 lifts per hand) of the test object with the baseline weight set to 3.2 N. On the next right-hand lift, the test object weight was increased to 14 N for a single catch trial. Throughout the experiment, participants held a 400 g (3.92 N) disk in their right hand (Fig. 1A) but could still use the tips of the index finger and thumb of the right hand to grasp and lift the test object. Therefore, in baseline trials, the total weight being supported by the right hand at the time of test object liftoff was 7.12 N. In other words, the joint torques required for right-hand lifts of the 3.2 N test object in experiment 3 were similar to those required to the lift of the 7 N object in experiment 2. Note we had participants perform alternating right- and left-hand lifts so that in baseline trials they received strong contextual information distinguishing the unchanging weight of the test object (3.2 N) from the weight of the disk held in the right hand.
Figure 5B shows the VF trace from the catch trial for every participant in the Control condition. We have added 3.92 N to the VF recorded from the test object so that the VF shown in the figure corresponds to the total VF supported by the hand. As illustrated in Figure 5, A and B, the slope of the initial corrective response in the Control condition was smaller than the slope in the 7-20 condition and similar to the slope in the 3-20 condition. Independent samples t tests showed that the corrective VF slope (over the first 200 ms of the corrective response) in the Control condition was smaller (t(17) = 3.29, p = 0.004) than in the 7-20 condition but did not differ (t(19) = 0.57, p = 0.579) from the 3-20 condition (Fig. 5C). These results indicate that the increase in the gain of the corrective lifting response with baseline weight observed in experiment 2 is not because of peripheral gain scaling.
Note that participants in the Control condition lifted the 3.2 N test object in roughly the same amount of time as participants in experiment 2. Therefore, we would expect the initial rate of VF increase during the predictive phase of the lift in the Control condition to be similar to the rate in the 3-20 condition and smaller than the rate in the 7-20 condition. In line with these expectations, independent samples t tests showed that the VF rate at half the weight of the test object, averaged across the 13 baseline lifts with the right hand, in the Control condition was smaller (t(17) = 4.97, p < 0.001) than in the 7-20 condition but did not differ (t(19) = 1.76, p = 0.095) from the 3-20 condition (Fig. 5D).
Discussion
The aim of this study was to test the hypothesis that the gain of the corrective lifting response, initiated when lifting an object that is heavier than expected, can be intelligently adapted to different contexts. In experiment 1, participants repeatedly lifted a 2 N object, which on occasional catch trials (i.e., lifts) increased to a weight that in separate experimental sessions was set to either 4, 6, or 10 N. We found that the gain of corrective response quickly scaled to the greater weight within a few catch trials. In experiment 2, different groups of participants repeatedly lifted an object with a baseline weight of either 3 or 7 N. The weight was then unexpectedly increased, and we found that the gain of the corrective response scaled with the baseline weight. Finally, in experiment 3, we provide evidence that the scaling of the gain of the corrective response with baseline weight observed in experiment 2 is not because of peripheral, automatic gain scaling linked to the larger joint torques required to lift the 7 N object compared with the 3 N object.
Research on corrective responses supporting goal-directed behavior has focused on responses elicited by the presence of unexpected sensory signals—signals that provide detailed information about the perturbation (e.g., the force applied to the hand). In contrast, the corrective lifting response, triggered by the absence of expected sensory signals, can be viewed as belonging to a different class of response that is initiated without complete sensory information about the perturbation. When the lifting response is triggered, the motor system only has categorical information about the error; it knows that the object is heavier than expected but not its actual weight. Therefore, any scaling of the response must be based solely on past experience. (In contrast, responses elicited by the presence of unexpected sensory signals can be shaped by both present and past experience; Crevecoeur and Scott, 2013.) Of course, once the object actually lifts off, its actual weight can be sensed, and information about the perturbation—and hence the sensory prediction error—is available. Although the sensory prediction error cannot be used to shape the current response, our results show it can be used to update the gain of the response in memory. Importantly, this class of corrective response may form part of control policies implemented in any task in which we predictively apply a target force to produce a desired outcome. If the target force is less than the required force, the desired effect will not occur at the predicted time, resulting in the absence of expected sensory signals triggering a corrective increase in the applied force. Another scenario in which the size of the perturbation is unknown when a sensory prediction error occurs is when moving to contact an unseen target that is farther away than expected. We are unaware of any reaching studies that have examined corrective responses in this scenario, but responses to unexpected changes in support surface height have been examined in walking. When the surface is lower than expected, the absence of predicted sensory feedback related to heel contact elicits functionally relevant leg muscle responses within ∼50–100 ms (van der Linden et al., 2007; van Dieën et al., 2007; Shinya et al., 2009). Note that unlike the corrective lifting response, which is tailored to achieving the primary task goal (i.e., liftoff), these locomotor responses are geared toward maintaining stability. Our task differs from other tasks in which a series of exploratory movements are made to locate a target (Stafford et al., 2012; Bollu et al., 2021). In this case, a given exploratory movement may result in failure (i.e., an absence of sensory feedback indicating that the target has been contacted), but this leads to further exploratory movements. In contrast, in the lifting response, the initial movement, which is expected to succeed, is followed by a very distinct automatic corrective response.
Numerous studies have shown that the gain of corrective responses in target-directed reaching can be tailored to different contexts to ensure that the movement goal is achieved while avoiding unnecessary energy expenditure, a key feature of optimal feedback control referred to as the principle of minimum intervention (Todorov and Jordan, 2002). For example, when the actual or viewed position of the hand is perturbed sideways while reaching straight ahead to a target, the gain of the corrective response (moving the hand back toward the target) scales, inversely, with the width of the target (Saunders and Knill, 2004; Knill et al., 2011; Nashed et al., 2012; Dimitriou et al., 2013; Diamond et al., 2015; Gallivan et al., 2016; Franklin et al., 2017). Importantly, provided the perturbation is not too large, these corrective responses tend to preserve the efficacy of the response such that the motor system does not have to plan and execute a second voluntary movement, at a considerable reaction time cost, to attain the goal (Goodale et al., 1986; Desmurget et al., 2004; Franklin and Wolpert, 2008). Thus, the question arises whether the gain changes in the corrective lifting response preserve its efficacy across contexts. To address this question, we can consider how quickly this voluntary strategy would achieve object liftoff in comparison with the corrective response. Generating a second volitional increase in vertical force would likely involve a reaction time of ∼300 ms (Danion and Sarlegna, 2007) and take ∼150 ms to produce, assuming that the first and second voluntary responses would be similar (Figs. 2, 4). Thus, the time from the initial predicted liftoff time, when a sensory prediction error occurs in catch trials, to liftoff would be ∼450 ms. This is considerably longer than the total corrective response durations (from predicted to actual liftoff) observed in the 2-4 and 2-6 conditions and even longer than the total corrective response duration observed in the 2-10 condition (∼380 ms), in which the vertical force must increase by 8 N during the corrective response (Fig. 3D). Had the gain of the corrective response (i.e., the corrective VF slope) observed in the 2-4 condition (∼11 N/s) not increased as the catch weight increased, the total corrective response direction in the 2-10 would have been ∼807 ms (80 ms + 8 N / 0.011 N/ms). Thus, the adaptability of the gain of the corrective lifting response preserves the efficacy of the response. The scaling of the gain of the corrective response to the baseline weight observed in experiment 2 likely also tends to preserve the efficacy of the response in real-world scenarios. On the assumption that the variation in weight across similar objects is proportional to the mean object weight, we would expect that when an object is heavier, on average, the weight perturbation (actual minus predicted weight) would be greater. Thus, scaling the gain of the corrective response to the baseline weight would compensate for the increase in the expected perturbation, helping to preserve the efficacy of the response across contexts.
When an object is lifted, the breaking of contact between the object and surface is signaled by fast-adapting Type II mechanoreceptors that are extremely sensitive to mechanical transients and provide robust and temporally precise feedback about liftoff (Johansson and Cole, 1992; Johansson and Edin, 1993). Thus, by comparing the actual and predicted tactile signals, the motor system can precisely monitor whether liftoff occurs when expected. However, the motor system could also make use of other sensory signals (including visual, proprioceptive, and even auditory signals) to monitor liftoff (Gale et al., 2021a,b). Accurate prediction of sensory outcomes when lifting requires knowledge of object weight. There is evidence that information about object weight during the performance of lifting tasks is represented in both dorsal and ventral visual areas and motor and premotor areas (Chouinard et al., 2005; Jenmalm et al., 2006; van Nuenen et al., 2012; Gallivan et al., 2014). When weight predictions are erroneous, the right inferior parietal cortex appears to be involved in detecting the mismatch between the actual and predicted weight (Jenmalm et al., 2006). During the corrective response that follows an underestimation of object weight, activity increases in the primary motor and somatosensory cortices and decreases in the cerebellum, with the opposite pattern of activity observed when the weight is overestimated (Jenmalm et al., 2006).
The task of grasping and lifting an object represents a rich paradigm for examining the interaction between predictive and reactive sensorimotor motor processes because it engages multiple corrective responses that function both to preserve grasp stability and achieve the task goal. Thus, when the digits first contact the object, mismatches between actual and predicted tactile information about the contact interface—including its angle, shape, and slipperiness—lead to updating of the coordination of fingertip forces within ∼90 ms. (Johansson and Westling, 1984, 1987; Jenmalm and Johansson, 1997; Jenmalm et al., 2000). Accidental slips at the contact interface result in similarly rapid updating of force coordination that is sensitive to the phase of the task. Thus, if a slip occurs while vertical forces are being generated before liftoff, the corrective response involves decreasing vertical force; whereas, if a slip occurs while the object is being held aloft, the corrective response involves increasing horizontal grip force (Johansson and Westling, 1984). Finally, different corrective responses are generated if an object is heavier or lighter than expected; in the latter case the response involves terminating the increase in vertical force (as well as grip force) and repositioning the object (Johansson and Westling, 1988). Thus, the simple task of lifting an object involves a sophisticated control policy governing how sensory feedback is used in real time to generate motor commands. Here, we have shown that the gain of one of the most commonly evoked corrective responses governed by this policy can be flexibly adapted based on short- and long-term experience.
Footnotes
This work was supported by Canadian Institutes for Health Research Grant RGPIN/05944-2019 to J.R.F., Swedish Research Council Project Grant 22209 to R.S.J., Natural Science and Engineering Research Council of Canada Grant 156173 (to J.R.F.), and National Institutes of Health Grants R01NS117699 and U19NS104649 to D.M.W. We thank Sean Hickman, Martin York, and Lee Baugh for technical support and Justin Caldwell for help with recruiting participants.
The authors declare no competing financial interests.
- Correspondence should be addressed to J. Randall Flanagan at flanagan{at}queensu.ca
