Internal Models Underlying Predictive Force Control

As illustrated in Fig. 3A, with objects of different weight, people use different rates of force increase prior to liftoff. Since there is no sensory information available about object weight until liftoff, this behavior indicates that people predict the final force requirements. Likewise, with objects of different friction (Fig. 3B) and shape (Fig. 3C), the force output is tailored to the properties of the object from the start of the initial force attack, well before sensory information from the digits obtained after contact with the object could have exerted any influence. Thus, in all three cases, the motor controller operates in a feedforward fashion and uses motor command parameters determined by internal models that capture the physical properties of the object. Figure 4 further illustrates that such internal models also capture dynamic properties of objects. The question arises as to how such models are selected and updated for different objects and after changes in object properties.

1. Prediction Based on Object Shape

Figures 5A and 5B show three consecutive trials taken from a series of lifts in which the angle of the grasped surfaces was changed between trials in a pseudorandom order. The sequence is 30°, —30°, and —30° and thus includes a transition from an upward tapered object (30°) to a downward tapered object (—30°). In the trials preceding this sequence, a 30° object was lifted. First consider the trials in which vision of the objects is available (Fig. 5A). When the shape of the object is changed, the grip force is adjusted from the very start of the lift in anticipation of the lower grip force required to lift the object. In particular, grip force is now increased more slowly before sensory feedback from the digits could have influenced the motor output. The predictive adjustment in grip force observed in the first trial after the switch in object shape is very accurate. Indeed, no further adjustment is

Figure 5 (A and B) Force adjustments to changes in surface angle during lift series in which surface angle was unpredictably varied between lift trials. Vertical load force, horizontal grip force, and grip force rate shown as a function of time for trials with (A) and without (B) vision and with normal digital sensibility. The dotted curves are from the last trial before the switch with the 30° object. The solid curves show the next trial with the —30° object. These curves illustrate adjustments to the smaller angle. The dashed lines show the following trial again with the —30° object. The downward arrow in B indicates the point in time when the new surface angle was expressed in terms of motor output. (C and D) Adaptation to surface shape during digital anesthesia with (C) and without (D) vision. Vertical load force, horizontal grip force, and grip force rate as a function of time for trials with 30° (dotted lines) 0° (solid lines and —30°, (dashed lines) surface angle (modified with permission from Jenmalm, P., and Johansson, R. S., J. Neurosci. 17, 4486-4499, 1997. Copyright © 1997 by the Society of Neuroscience).

Figure 5 (A and B) Force adjustments to changes in surface angle during lift series in which surface angle was unpredictably varied between lift trials. Vertical load force, horizontal grip force, and grip force rate shown as a function of time for trials with (A) and without (B) vision and with normal digital sensibility. The dotted curves are from the last trial before the switch with the 30° object. The solid curves show the next trial with the —30° object. These curves illustrate adjustments to the smaller angle. The dashed lines show the following trial again with the —30° object. The downward arrow in B indicates the point in time when the new surface angle was expressed in terms of motor output. (C and D) Adaptation to surface shape during digital anesthesia with (C) and without (D) vision. Vertical load force, horizontal grip force, and grip force rate as a function of time for trials with 30° (dotted lines) 0° (solid lines and —30°, (dashed lines) surface angle (modified with permission from Jenmalm, P., and Johansson, R. S., J. Neurosci. 17, 4486-4499, 1997. Copyright © 1997 by the Society of Neuroscience).

observed on the second trial after the change when information about shape has been obtained through tactile sensory signals. These results demonstrate that visual geometric cues can be used to efficiently specify the force coordination for object shape in a feedforward manner. These cues are used to parametrically adapt the finger force coordination to object shape in anticipation of the upcoming force requirements.

When vision of the object is not available, a very different pattern of force output is obtained. On the first trial after the switch to the —30° object, grip force develops initially according to the force requirements in the previous trial. This indicates that memory of the previous surface angle determines the default force coordination in a feedforward manner. However, about 100 msec after the digits contacted the object, the grip force was modified and tuned appropriately for the actual surface angle (see first trial with the —30° in Fig. 5B). This amount of time is required to translate tactile information into motor commands, a process that likely involves supraspinal processing. By the second trial after the switch, the force output is appropriately adapted to the —30° surface angle from the onset of force application. Thus, an internal model related to object shape determines the force coordination in a feedforward fashion and tactile sensory information obtained at initial contact with the object mediates an updating of this model to changes in object shape. Furthermore, a single trial is enough to update the relevant internal model.

Sensors in the digits are thus used to update the force coordination for object shape when visual cues are unavailable or misleading. When digital sensibility is removed by local anesthesia, leaving neither visual nor somatosensory cues about shape, the adaptation in force output is severely impaired (Fig. 5D). Although grip force and load force still change in parallel, force output is no longer updated following contact. People adapt to the loss of both visual and tactile sensory cues about shape by applying strong grip forces regardless of surface angle. When vision is available during digital anesthesia, people are able to adapt their forces to object shape with only minor impairments (Fig. 5C). Thus, visual geometric cues can be used effectively for feedforward control even in the absence of somato-sensory cues about shape.

The curvature of the grasp surfaces is another aspect of object shape. Surprisingly, the curvature of spherically curved symmetrical grasp surfaces has little effect on grip force requirements for grasp stability under linear force loads. However, it becomes acute in tasks involving torsional loads. The relationship between the grip force and tangential torque is parametrically scaled by surface curvature: For a given torque load, people increase grip force when curvature increases. As with linear force loads, this scaling of grip force is directly related to the minimum grip force required to prevent slip. Under torsional loads, people maintain a small but adequate safety margin against rotational slip. As with surface angle, visual information about surface curvature can be used for feedforward control of force. Likewise, people use cues provided by tactile afferents to adapt force once finger contact is established.

2. Prediction Based on Object Weight

When we manipulate familiar or common objects that we can identify either visually or haptically, we are extremely adept at selecting fingertip forces that are appropriately scaled to the weight of the object. That is, during the very first lift of a common object, before sensory information related to weight becomes available at liftoff, the force development is tailored to the weight of the object. This indicates that we can use visual and haptic cues to select internal models that we have acquired for familiar objects and can use these models to parametrically adjust our force output to object weight. For "families" of familiar objects that vary in size (e.g., screwdrivers, cups, soda cans, and loafs of bread), we can exploit size-weight associations, in addition to object identity, to scale our force output in a feedforward fashion. However, as we have all experienced, our force output may sometimes be erroneous. Such situations can be created experimentally by unexpectedly changing the weight of a repeatedly lifted object without changing its visual appearance. In such cases, the lifting movement may be either jerky or slow. For example, if the object is lighter than expected from previous lifting trials, the load force and grip force drives will be too strong when the load force overcomes the force of gravity and liftoff takes place. Although somatosensory afferent events, evoked by the unexpectedly early liftoff, trigger an abrupt termination of the force drive, this occurs too late (due to control loop delays) to avoid an excessively high lift. Burst responses in FA II (Pacinian) afferents, which show an exquisite sensitivity to mechanical transients, most quickly and reliably signal the moment of liftoff. Conversely, if the object is heavier than expected, people will initially increase load force to a level that is not sufficient to produce liftoff and no sensory event will be evoked to confirm liftoff (Fig. 6A, solid curves). Importantly, this absence of a sensory event at the expected liftoff causes the release of a new set of motor commands. These generate a slow, discontinuous force increase until terminated by a neural event at the true liftoff (Fig. 6A, afferent response during the 800-g lift following the 400-g lift). Taken together, these observations indicate that control actions are taken as soon there is a mismatch between an expected sensory event and the actual sensory input. Thus, the absence of an expected sensory event may be as efficient as the occurrence of an unexpected sensory event in triggering compensatory motor commands. Moreover, this mismatch theory implies that somatosensory signals that represent the moment of liftoff are mandatory for the control of the force output whether or not the weight of the object is correctly anticipated. Finally, once an error occurs, the internal model of the object is updated to capture the new weight. In natural situations, this generally occurs in a single trial. As shown in Fig. 6A, in the trial after the switch trials when the weight of the object was unexpectedly increased from 400 to 800 g, the forces were correctly scaled for the greater weight (dashed curves).

3. Prediction Based on Friction

Whereas people use visual information about object size and shape to scale fingertip forces, there is no

Figure 6 Single unit tactile afferent responses and adjustments in force to changes to object weight (A) and to the frictional condition between the object and the digits (B). Data are from single lift trials. (A) Three successive trials in which the subject lifted a 400-g object (dotted curves), an 800-g object (solid curves), and then the 800-g object again (dashed curves). The forces exerted in the first lift are adequately programmed because the subject had previously lifted the 400-g object. The forces are erroneously programmed in first lift of the 800-g object because they are tailored for the lighter 400-g object lifted in the previous trial. The vertical lines with arrowheads pointing downward indicate the moment of liftoff for each trial and they indicate the evoked sensory events exemplified by signals in a single FA II afferent. The absence of burst responses in FA II afferents at the expected point in time for the erroneously programmed 800-g trial is used to initiate a new control mode. This involves slow, discontinuous, and parallel increases in grip force and load force until terminated by sensory input signaling liftoff. (B) The influence of friction on force output and initial contact responses in a FA I unit. Two trials are superimposed, one with less slippery sandpaper (dashed lines) and a subsequent trial with more slippery silk (solid lines). The sandpaper trial was preceded by a trial with sandpaper and therefore the force coordination is initially set for the higher friction. The vertical line indicates initial touch (modified with permission from Johansson, R. S., and Westling, G., Exp. Brain Res. 66,141-154,1987. Copyright © 1987 by Springer-Verlag; and from Curr. Opin. Neurobiol. Johansson, R. S., and Cole, K. J., 2, 815-823, Copyright © 1992, with permission from Elsevier Science).

Figure 6 Single unit tactile afferent responses and adjustments in force to changes to object weight (A) and to the frictional condition between the object and the digits (B). Data are from single lift trials. (A) Three successive trials in which the subject lifted a 400-g object (dotted curves), an 800-g object (solid curves), and then the 800-g object again (dashed curves). The forces exerted in the first lift are adequately programmed because the subject had previously lifted the 400-g object. The forces are erroneously programmed in first lift of the 800-g object because they are tailored for the lighter 400-g object lifted in the previous trial. The vertical lines with arrowheads pointing downward indicate the moment of liftoff for each trial and they indicate the evoked sensory events exemplified by signals in a single FA II afferent. The absence of burst responses in FA II afferents at the expected point in time for the erroneously programmed 800-g trial is used to initiate a new control mode. This involves slow, discontinuous, and parallel increases in grip force and load force until terminated by sensory input signaling liftoff. (B) The influence of friction on force output and initial contact responses in a FA I unit. Two trials are superimposed, one with less slippery sandpaper (dashed lines) and a subsequent trial with more slippery silk (solid lines). The sandpaper trial was preceded by a trial with sandpaper and therefore the force coordination is initially set for the higher friction. The vertical line indicates initial touch (modified with permission from Johansson, R. S., and Westling, G., Exp. Brain Res. 66,141-154,1987. Copyright © 1987 by Springer-Verlag; and from Curr. Opin. Neurobiol. Johansson, R. S., and Cole, K. J., 2, 815-823, Copyright © 1992, with permission from Elsevier Science).

evidence that they use visual cues to control the balance of grip and load force for friction. However, tactile receptors in the fingertips are of crucial importance. The most important adjustment after a change in friction takes place shortly after the initial contact with the object and can be observed about 100 msec after contact (Fig. 6B). Prior to this force adjustment, there are burst responses in tactile afferents of different types but most reliably in the population of FA I (Meissner) afferents. The initial contact responses in subpopulations of excited FA I afferents are markedly influenced by the surface material as exemplified in Fig. 6B with a single afferent. The adjustment of force coordination to a change in frictional condition is based on the detection of a mismatch between the actual and an expected sensory event. This adjustment involves either an increase in the grip-to-load force ratio if the surface is more slippery than expected (as shown in Fig. 6B) or a decrease in the ratio of the surface if less slippery than expected. The adjustment also includes an updating of the internal model so as to capture the new frictional conditions between the object and the skin for predictive control of the grip-to-load force ratio in further interactions with the object. However, sometimes these initial adjustments to frictional changes are inadequate and an accidental slip occurs at a later point, often at one digit only. Burst responses in dynamically sensitive tactile afferents to such slip events promptly trigger an automatic upgrading of the grip-to-load force ratio to a higher maintained level. This restores the grip force safety margin during subsequent manipulation by updating the internal model controlling the balance between grip and load force.

In summary, skilled manipulation involves two major types of control processes: anticipatory parameter control and discrete event, sensory-driven control. Anticipatory parameter control refers to the use of visual and somatosensory inputs, in conjunction with internal models, to tailor finger tip forces for the properties of the object to be manipulated prior to the execution of the motor commands. For familiar objects, visual and haptic information can be used to identify and select the appropriate internal model that is used to parametrically adapt motor commands, prior to their execution, in anticipation of the upcoming force requirements. People may also use geometric information (e.g., size and shape) for anticipatory control, relying on internal forward models capturing relationships between geometry and force requirements. There is ample evidence that the motor system makes use of internal models of limb mechanics, environmental objects, and task properties to adapt motor commands.

Discrete event, sensory-driven control refers to the use of somatosensory information to acquire, maintain, and update internal models related to object properties. This type of control is based on the comparison of actual somatosensory inflow and the predicted somatosensory inflow—an internal sensory signal referred to as corollary discharge. (The soma-tosensory input provided by tactile signals in the digital nerves is obviously critical in the control of skillful manipulation.) Thus, when we lift an object, we generate both efferent motor commands to accomplish the task and this internal sensory signal. Together, these are referred to as the sensorimotor program. Predicted sensory outcomes are produced by an internal forward model in conjunction with a copy of the motor command (referred to as an efference copy). Disturbances in task execution due to erroneous parameter specification of the sensorimotor program give rise to a mismatch between predicted and actual sensory input. For example, discrete somatosensory events may occur when not expected or may not occur when they are expected (Fig. 6A). Detection of such a mismatch triggers preprogrammed patterns of corrective responses along with an updating of the relevant internal models used to predict sensory events and estimate the motor commands required. This updating typically takes place within a single trial. With respect to friction and aspects of object shape, the updating primarily occurs during the initial contact with the object. In trials erroneously programmed for object weight and mass distribution, the updating takes place when the object starts to move (e.g., at liftoff in a lifting task).

Was this article helpful?

0 0
Breaking Bulimia

Breaking Bulimia

We have all been there: turning to the refrigerator if feeling lonely or bored or indulging in seconds or thirds if strained. But if you suffer from bulimia, the from time to time urge to overeat is more like an obsession.

Get My Free Ebook


Post a comment