浙江大学 文本 Forward modeling allows feedback control for fast reaching movements_生理学

Decmurget and Grattan -Farward madallng Review Forward modeling allows feedback control for fast reaching movements Michel Desmurget and Scott Grafton ays in otor loops have led to the al that re ching mov ments an ery end of a traje ry.The ut is updated c sly by int e of the mo s sem In such ar ty of an effec ed with ne f ng mo rietal lobe and cerel rate for mental operations that reguire an estimateof uelae in the immediate future. 164 “eaen0 nv“0,2808

Since the early contribution of Woodworth1 , the degree to which visually-directed movements are planned in ad￾vance or controlled online during their actual execution has been an issue of considerable debate2–6. After almost a century of controversy, the relative importance of three dif￾ferent models, namely the feedforward, feedback and hy￾brid, continues to be argued. Feedforward models propose that a motor command is defined in advance of the onset of 423 Forward modeling allows feedback control for fast reaching movements Michel Desmurget and Scott Grafton Delays in sensorimotor loops have led to the proposal that reaching movements are primarily under pre-programmed control and that sensory feedback loops exert an influence only at the very end of a trajectory. The present review challenges this view. Although behavioral data suggest that a motor plan is assembled prior to the onset of movement, more recent studies have indicated that this initial plan does not unfold unaltered, but is updated continuously by internal feedback loops. These loops rely on a forward model that integrates the sensory inflow and motor outflow to evaluate the consequence of the motor commands sent to a limb, such as the arm. In such a model, the probable position and velocity of an effector can be estimated with negligible delays and even predicted in advance, thus making feedback strategies possible for fast reaching movements. The parietal lobe and cerebellum appear to play a crucial role in this process. The ability of the motor system to estimate the future state of the limb might be an evolutionary substrate for mental operations that require an estimate of sequelae in the immediate future. M. Desmurget is at INSERM U534, ‘Space and Action’, 16 av. du Doyen Lépine, 69500, Bron, France. S. Grafton is at the Center for Cognitive Neuroscience, 6162 Moore Hall, Dartmouth College, Hanover, NH 03755, USA. tel: +1 603 646 0038 fax: +1 603 646 1181 e-mail : Scott.T.Grafton@ dartmouth.edu Desmurget and Grafton – Forward modeling Review 1364-6613/00/$ – see front matter © 2000 Elsevier Science Ltd. All rights reserved. PII: S1364-6613(00)01537-0 Trends in Cognitive Sciences – Vol. 4, No. 11, November 2000 15 Miller, G.A. (1956) The magical number seven, plus or minus two: some limits on our capacity for processing information. Psychol. Rev. 63, 81–97 16 Vallar, G. and Baddeley, A.D. (1984) Fractionation of working memory: neuropsychological evidence for a phonological short–term store. J. Verbal Learn. Verbal Behav. 23, 151–161 17 Wilson, B.A. and Baddeley, A.D. (1988) Semantic, episodic and auto￾biographical memory in a post–meningitic amnesic patient. Brain Cognit. 8, 31–46 18 Baddeley, A.D. Levels of working memory. In Perspectives on Human Memory and Cognitive Aging: Essays in Honor of Fergus Craik (Naveh Benjamin, M. et al., eds), Psychology Press (in press) 19 Ericsson, K.A. and Kintsch, W. (1995) Long-term working memory. Psychol. Rev. 102, 211–245 20 Bartlett, F.C. (1932) Remembering, Cambridge University Press 21 Schank, R.C. (1982) Dynamic memory, Cambridge University Press 22 Neath, I. and Nairne, J.S. (1995) Word-length effects in immediate memory: overwriting trace-decay theory. Psychonomic Bull. Rev. 2, 429–441 23 Gathercole, S.E. and Hitch, G.J. (1993) Developmental changes in short￾term memory: a revised working memory perspective. In Theories of Memory (Collins, A. et al., eds), pp. 189–209, Lawrence Erlbaum 24 Hickok, G. and Poeppel, D. (2000) Towards a functional neuroanatomy of speech perception. Trends Cognit. Sci. 4, 131–138 25 Baddeley, A.D. and Logie, R.H. (1992) Auditory imagery and working memory. In Auditory Imagery (Reisberg, D., ed.), pp. 179–197, Lawrence Erlbaum 26 Baddeley, A.D. and Andrade, J. (2000) Working memory and the vividness of imagery. J. Exp. Psychol. Gen. 129, 126–145 27 Baddeley, A.D. (1993) Working memory and conscious awareness. In Theories of Memory (Collins, A.F. et al., eds), pp. 11–28, Lawrence Erlbaum 28 Tulving, E. (1989) Memory: performance, knowledge and experience. European J. of Cog. Psychol. 1, 3–26 29 Hummel, J. (1999) The binding problem. In The MIT Encyclopedia of the Cognitive Sciences (Wilson, R.A. and Keil, F.C., eds), pp. 85–86, MIT Press 30 Singer, W. (1999) Binding by neural synchrony. In The MIT Encyclopedia of the Cognitive Sciences (Wilson, R.A. and Keil, F.C., eds), pp. 81–84, MIT Press 31 Roberts, A.C. et al. (1996) Executive and cognitive functions of the prefrontal cortex. Philos. Trans. R. Soc. London Ser. B 351, 1387–1527 32 Prabhakaran, V. et al. (2000) Integration of diverse information in working memory within the frontal lobe. Nat. Neurosci. 3, 85–90 33 Cowan, N. (1999) An embedded-processes model of working memory. In Models of Working Memory (Miyake, A. and Shah, P., eds), pp. 62–101, Cambridge University Press 34 Jones, D.M. (1993) Objects, streams and threads of auditory attention. In Attention: Selection, Awareness and Control (Baddeley, A.D. and Weiskrantz, L., eds), pp. 87–104, Clarendon Press 35 Allport, D.A. (1984) Auditory–verbal short-term memory and conduction aphasia. In Attention and Performance Vol. X: Control of Language Processes (Bouma, H. and Bouwhuis, D.G., eds), pp. 313–326, Lawrence Erlbaum 36 Conway, M.A. et al. (1996) Recollections of true and false autobiographical memories. J. Exp. Psychol. Gen. 125, 69–95

Review @ in sia-vel.4.No.11.Nevember 3000

movement. Within this context, the role of feedback loops is, at most, marginal and circumscribed to the very end of the trajectory3,4,7–9 when hand velocity is low. Feedback models (Box 1), when regarded as the conceptual opposite of feedforward models, propose that the pattern of muscle activation that is required to point to the target is not defined prior to the onset of movement, but rather during the course of arm displacement. Thus, there is no a priori motor plan and the muscle command is generated in real time through an error signal that continuously compares the relative locations of the hand and target10,11. Hybrid models represent a trade-off between the feed￾forward and feedback hypotheses. In a hybrid model, a crude motor plan is assembled prior to the onset of movement (feedforward component). This initial plan does not unfold unattended, because it is imprecise12–15. Rather, it remains under the constant ‘supervision’ of powerful internal feed￾back loops that adjust and refine it in real time (feedback component). In this paper, we examine the validity of these three models for reaching movements of the hand in light of recent psychophysical, lesion-based and functional imaging studies. We show that both the feedforward and feedback hypotheses, when considered in isolation, are overly reduc￾tionistic. Consequently, we propose an integrative hybrid model of motor control in which preplanning and feedback control are both used by the nervous system. We first review the evidence that is generally believed to have established a dominant role for feedforward movement control, namely that sensory feedback loops are too slow to allow efficient trajectory control. Second, we show that feedback mech￾anisms can rely on much more than sensory inflow than has been thought traditionally. Feedback control strategies be￾come viable if the instantaneous location of the hand can be inferred by the nervous system through a forward model that integrates efferent and afferent signals to infer, with no delay, the current state of the motor system (Box 2). Third, we argue against the plausibility that purely feedforward strategies generate rapid reaching movements. To this end, we show that online control by visual and non-visual infor￾mation occurs early in a hand movement. We also provide evidence that the motor command is not generated exclu￾sively in real time, which is contrary to the suggestion of purely feedback models. Finally, we discuss briefly the func￾tional anatomy of internal feedback loops with emphasis on the potential contribution of the posterior parietal cortex (PPC) and the cerebellum. Review Desmurget and Grafton – Forward modeling 424 Trends in Cognitive Sciences – Vol. 4, No. 11, November 2000 The term ‘feedback’ has multiple meanings in the literature. In this review, it refers to processes that mediate hand path cor￾rections by comparing the target position and an estimation of the hand location. This is the primary feedback mechanism that allows modulation of the initial motor command when it is inaccurate. From a theoretical point of view, we can segre￾gate feedback loops into three categories: (1) sensory feedback, in which the location of the hand is estimated on the basis of sensory information; (2) non-sensory feedback, in which the location of the hand is estimated on the basis of efferent infor￾mation; and (3) internal feedback, in which the location of the hand is estimated on the basis of both efferent and afferent signals. The term feedback alone refers to any of these loops. Box 1. Different types of feedback Many of the concepts that are used to describe the planning and execution of human arm movements have been borrowed from robotic manipulators. Among these concepts, three are espe￾cially relevant for the present review, namely feedforward and feedback control and internal models. In a feedforward control system, the set of muscle activations that drives a reach towards a target is defined prior to arm dis￾placement. After the onset of movement, the motor command unfolds unaltered until the movement is completed. Computer programs that specify a series of commands to be performed in advance are a good analogy of feedforward control mechanisms. In a feedback control system, the current state of the system is compared to reference values. In the case of a discrepancy between these two parameters, an error signal is generated and used to modulate the behavior of the controlled system. The most common example of a feedback system is the thermostat, which compares the current temperature to a reference value and then modulates the response of the radiator. For reaching movements, the controlled system could be the arm, the refer￾ence state the target position and the current state the location of the hand. As long as the hand has not reached the target, a motor command is generated in real time (Refs a,b). Thus, muscle activations do not have to be specified in advance. Internal models can be segregated into two categories, namely forward models and inverse models (Refs c,d). A forward model predicts the behavior of the motor system in response to a command and allows the CNS to estimate the current and future state of the effector immediately and without peripheral information. Forward models are particularly interesting in the context of feedback control systems. For instance, a forward model can produce an estimate of the movement end-point lo￾cation as output, which can be compared to the target lo-cation. In the case of a discrepancy, a corrective command can be gen￾erated. An inverse model takes into account the inertial and vis￾cous properties of the arm to estimate the motor command that will produce the desired displacement. Inverse models are there￾fore critical for feedforward control systems (Refs e,f). References a Hinton, G. (1984) Parallel computations for controlling an arm. J. Mot. Behav. 16, 171–194 b Flanagan, J.R. et al. (1993) Control of trajectory modifications in target-directed reaching. J. Mot. Behav. 25, 140–152 c Wolpert, D.M. et al. (1998) Internal models in the cerebellum. Trends in Cognitive Sciences 2, 338–347 d Kawato, M. (1999) Internal models for motor control and trajectory planning. Curr. Opin. Neurobiol. 9, 718–727 e Katayama, M. and Kawato, M. (1993) Virtual trajectory and stiffness ellipse during multijoint arm movement predicted by neural inverse models. Biol. Cybern. 69, 353–362 f Gomi, H. and Kawato, M. (1996) Equilibrium-point control hypothesis examined by measured arm stiffness during multijoint movement. Science 272, 117–120 Box 2. From robots to humans

Review Rapid balistic Slow terminal arm lransport 具r5Ie Desred Motor model displacement plan Molor command Hand Target location ES Intial condoons Sensory feedback module de桥elodity is lew,The ourent lecation时ehds4e操en与d4uYn制wy司 Sentory informatioa and focdback contrel reaching movements in typically 300-700 m.Oa the bais of Dwring the lst thme decades,foedfrwad mod of move these seslts was conduded that seasony feedhack bops mere coetrol have indiputably been the mou infuestial couk not be usod to control hind trajectary. This domirance wea hased on the theoeericall alerrative thar To reconcile the aboverentioned realts with the fact feedback lops shoald re女ouive山on semory infommation. hat poaldrected movements are more accute when pro. nmtien th was esplicidy fommalaed in Keck's po. prioceptive or visu indormaio present.a du modd of acering moph,follu.The ccrpt of mote pro- moor cnntm wa propoucd Accrding to th modd gram mighr be viewed s asr of mde commands char are structured befoee a movement soqunce begims,and dat al ponents.The firt is driven entirely by a mooor plan and en. lows the mtire soqumnce to be carried out uniafluenced by wanes rapid tarport of the hand to the vicinity of the ta peripheral foedback Fellowing thi definicon the cr pet.The scond depend on senry ferdback loop and butkn of seory informarion o movemenr cone was alls the very endofthe trajecy.when he widdy isvetipated.The type of revalt led to the cnu mowement velocity beceme low.Claically,thee coeree. ion tha sensery foedhack played a marginal role in move rioms are viewed as a sries of one or moee sub-moemens ment accuracy.Fint.somatic deallerenetion did not peevent hat are gened at ds tm inevso the bais sbjecs from natiey courate ovm in he etial eror sign The daal modd found song sup dark.Secand,some moemert of short duraticn cruld be eatration that viewing the hnd during the complned before the minimum dday required to peocess first half of che majectory did not improve mowemnent accu. sensory isforrution.Thisd online coeroctiom that were racy more thn when the hand wa never viiblIs ad. hated on senury feedhack could produce unrahle trajec. dition,tendan vibratio erm hed that altering oories for poinring tha was perfoered ar a high ar mediumn velociry.Among thesems the first is peebobly the when the wibeacion was applied ar the end of the majec. leat cmiing becane deafferentation eperinen have tory.We shal cxamine aterrative interpectation of thee produced incs eaSpocifically,while omed. obserrations afrer developing the idea of forwand models. ies have indicated that humns ar animals deprived ol pro oherbave hon that deafferemed aent of the body with repect o the emironreent subiects eshbit dramai mo impime and g erares the same rerinal stimularion.To ac Thetrneee the wcofcmory fooluck o contl mowemene trajectory is bated on the phyiclegical arioas von Hele and Mreliraede dcly th is inhereet in semsoeimoeor lops.If the procesing proporod tht a 'copy'of the mooor command wn stored of seor foin long with repe the dararion of smwhere in he bn and d oimerpe the percepel movement,the ponicion of the hand will change dramcicaly input.This conduion wa euended and pencralined,lad. by the rime the foedhack signal sramts o inflence the ongoing ngo点econcep平of the forward mode The ide mceor command,thus rendcring the imlmcrted coerocto behind this cocrpt is that the nervous ytem can peopre inapprpri Bebaviourd eperine he hwa dhat sively 'lears'to eim the behavior of the mot plan in the minimum dday needed foe a vimal ce propeincepeive reponae to a given cammand.By intngrating information signal to indumnce an ongoing movement is 80-100 ms nent condicions.motor ourflow Rch5间，whie tht for the duration of vintlhy-directed and i the prubable ponition and velocicy af the 69

Sensory information and feedback control During the last three decades, feedforward models of move￾ment control have indisputably been the most influential. This dominance was based on the theoretical alternative that feedback loops should rely exclusively on sensory information, an assumption that was explicitly formulated in Keele’s pio￾neering monograph, as follows. ‘The concept of motor pro￾gram might be viewed as a set of muscle commands that are structured before a movement sequence begins, and that al￾lows the entire sequence to be carried out uninfluenced by peripheral feedback’16. Following this definition, the contri￾bution of sensory information to movement control was widely investigated. Three types of results led to the conclu￾sion that sensory feedback played a marginal role in move￾ment accuracy. First, somatic deafferentation did not prevent subjects from executing relatively accurate movements in the dark. Second, some movements of short duration could be completed before the minimum delay required to process sensory information. Third, online corrections that were based on sensory feedback could produce unstable trajec￾tories for pointing that was performed at a high or medium velocity. Among these arguments, the first is probably the least convincing, because deafferentation experiments have produced inconsistent results. Specifically, while some stud￾ies have indicated that humans or animals deprived of pro￾prioception are able to perform relatively accurate move￾ments5,17,18, other experiments have shown that deafferented subjects exhibit dramatic motor impairments5,19,20. The strongest evidence against the use of sensory feedback to control movement trajectory is based on the physiological delay that is inherent in sensorimotor loops. If the processing of sensory information is long with respect to the duration of movement, the position of the hand will change dramatically by the time the feedback signal starts to influence the ongoing motor command, thus rendering the implemented correction inappropriate21,22. Behavioural experiments have shown that the minimum delay needed for a visual or proprioceptive signal to influence an ongoing movement is 80–100 ms (Refs 5,6), while that for the duration of visually-directed reaching movements is typically 300–700 ms. On the basis of these results, it was concluded that sensory feedback loops could not be used to control hand trajectory21,22. To reconcile the abovementioned results with the fact that goal-directed movements are more accurate when pro￾prioceptive or visual information is present, a dual model of motor control was proposed3,4,7–9. According to this model (Fig. 1), reaching movements are segmented into two com￾ponents. The first is driven entirely by a motor plan and en￾sures rapid transport of the hand to the vicinity of the tar￾get. The second depends on sensory feedback loops and allows corrections at the very end of the trajectory, when the movement velocity becomes low. Classically, these correc￾tions are viewed as a series of one or more sub-movements that are generated at discrete time intervals on the basis of a retinal error signal5–8,23. The dual model found strong sup￾port in the demonstration that viewing the hand during the first half of the trajectory did not improve movement accu￾racy more than when the hand was never visible24,25. In ad￾dition, tendon vibration experiments showed that altering proprioceptive feedback affected movement accuracy only when the vibration was applied at the end of the trajec￾tory26. We shall examine alternative interpretations of these observations after developing the idea of forward models. Feedback and the need for forward modeling A displacement of the body with respect to the environment and vice versa generates the same retinal stimulation. To ac￾count for the ability of the nervous system to discriminate between these two situations, von Holst and Mittelstaedt27 proposed that a ‘copy’ of the motor command was stored somewhere in the brain and used to interpret the perceptual input. This conclusion was extended and generalized, lead￾ing to the concept of the forward model14,22,28–31. The idea behind this concept is that the nervous system can progres￾sively ‘learn’ to estimate the behavior of the motor plan in response to a given command. By integrating information that is related to initial movement conditions, motor outflow and sensory inflow, the probable position and velocity of the Desmurget and Grafton – Forward modeling 425 Trends in Cognitive Sciences – Vol. 4, No. 11, November 2000 Review trends in Cognitive Sciences Desired displacement ES Initial conditions Sensory feedback module Rapid ballistic arm transport Slow terminal adjustment Sensor Inverse model Motor plan Hand location Corrective command Motor command Hand location Target location Fig. 1. The classical dual model of movement control. The required arm displacement is estimated based on the respective locations of the hand and target. This displacement is then converted into a motor plan through an inverse model. The main part of the movement un￾folds under the rigid control of this plan (ballistic arm transport). Sensory feedback loops become active at the very end of the movement, when velocity is low. The current location of the hand is then compared to the target position. In case of a discrepancy, an error signal (ES) is issued and a series of corrective sub-movements is generated. The movement stops when the hand reaches the target (circle in diamond)

Review D中nur身时+4g14相●n。手容4号的参眼ing Continuous control Desired rwer8e Molor model displacement plan Motor command Comective Motor outfow sensory inflow Hand Target Forward location ocston ES Final state esomale 果带华带。带带果票票带带华 Intitial conditions Feedback module Fig.2.Smplified hybrid circuits that makes ute of a forwaed model af arm dynamis for centrolling hand mevementt.A or plan b int与defined baned on The respective localiens of the hand and target Dur同eer成alorwerd model of the 钟时ete=sdp4T4特s0d特种a941nme6d4u%n effecor can be determsined (and even predicned).When a for- deaffcreataon of e propeooeption in moakeyM. ardmoddsd feed an inemfoedbadk lop,coml These results have,thus,sevealed the esisrence of a nonsen performance i impeoved ignificantly inamuch asge sory ferdback loop tht can accou fo the remarkable ddays tha aocd with sry foodback loops can be curay of the cdicem under noemal cnd avided.The advantas of forward moddling for mmere Far arm movemeats,the ecineece of non-scmory feed- conmrad shematically in Pig 2hen reqaired back loopse se of meorufwas inlys to reach a target,a ubjeet fint chboeno a mooor plan.bucd peed by behavioral studie that howed that hand raje on the inirial movemer codthe ep ca rory coud be amended with a shoner larency than the tioes of the hnd and target).Duting the sealication of the minim Lneacy toquired to process peripherl information. movement.a forwand model ef the dynamics of the am Fee imstance.Higgins and Angel oboerved that the pnered mlet venion,thi modd tios time o an unexpected modificarinn of the tanget cpy of the motor oufw.Based onisnmnthe majecrory in a manusl tracking task was shoner than the end-point of the movement can he prodicted and contin- proprioceptive reaction timne.A similar realt was arported uoudy comared to the targer lcation.Discrepancies cause by arwho found that altering the poprioceptive an ial tohe geerated.which riggena modalation tignal through vibracin did not modify the reaction time to of the motor commnd.More comples appeoacesavesu visual pemurbation.Using a poining rask.Cook and pored th forward modd do not anye motr oufow. Diggl oberved coerectios to the hand path within bar ake wse sen町nflou.This view is则pponed每 4 ms when the initial dieection of movemear was inoer. bchieral obcr(bdw)and hn been modeled by rect.This value was cose to thit teported by van Sonderen Mull ara with reference to an engincering control scheme eratm)ina doubleatep task in which the initid tar that is known as the Smith prodictne.In this acheme,forward nd during ar after the initieian of prdcis delyed by a peried that is cmarableo the cwo町，hu making it pouibie to comp Convincing arguments were tepurted in an clegant dicy the pedicd and srybademAny yp特代，hich d ha rliablees that tets from thi compatien can then be usod to update timatioan of the locarion of the hand could he chtained by the curent forward model of the dynamics of the amm. combining e and afferent sigakinward modd. One line af eridence that shows that non-ry feed- Thewe authon noquired wahjecto move their hand alonga back loops can be used to guide biclogia acmtors s found line while holing a manipuloor.The hand was allowed to with cye movements.Thete is now comiderable evidence be victod for 2 a prior to the onset of movement.The ma that the oculomor us gnao cn nipulr wascda mr that indcde tol accadic cye movemears Perturbarion experimea uive ar auistive force to the mouement.At the ed of the indicae thge s shified during the prep trial.the subjects escimated the location of their hand using cruion of a de oward a flabed target,then a cum a vial apot coned by the oher hand.The temporal pensry sccade which acorady brings gae on the propagarion of measured eroes eshibited by the sabjecrs remembered target location i grnerated (head-fised sac. coul be fuly accounoed foe by amuming that the motoe cade and bead-frce ge shifu).Strikingly,sch compen. coatrel syatem inegrates both motor oucflw and semoery in does not reuire vi propriocpeive fodhack, inlwo eim the eion of the hand.By c because it can eccur in complere darkness and after surgical modes based edusively on cicher sensory inflow o moto end:le《ogniniee sciences=VaL,4.ne,i1,Nee◆nb#上aao

effector can be determined (and even predicted). When a for￾ward model is used to feed an internal feedback loop, control performance is improved significantly inasmuch as large delays that are associated with sensory feedback loops can be avoided. The advantage of forward modeling for movement control is illustrated schematically in Fig. 2. When required to reach a target, a subject first elaborates a motor plan, based on the initial movement conditions (i.e. the respective loca￾tions of the hand and target). During the realization of the movement, a forward model of the dynamics of the arm is generated. In its simplest version, this model receives as input a copy of the motor outflow. Based on this information, the end-point of the movement can be predicted and contin￾uously compared to the target location. Discrepancies cause an error signal to be generated, which triggers a modulation of the motor command. More complex approaches have sug￾gested that forward models do not only use motor outflow, but also use sensory inflow29,30. This view is supported by behavioral observations (see below) and has been modeled by Miall et al.28 with reference to an engineering control scheme that is known as the Smith predictor. In this scheme, forward prediction is delayed by a period that is comparable to the sensory delay, thus making it possible to compare directly the predicted and sensory-based estimates. Any error that results from this comparison can then be used to update the current forward model of the dynamics of the arm. One line of evidence that shows that non-sensory feed￾back loops can be used to guide biological actuators is found with eye movements. There is now considerable evidence that the oculomotor system uses an efferent signal to con￾trol saccadic eye movements. Perturbation experiments indicate that if gaze is shifted during the preparation or ex￾ecution of a saccade toward a flashed target, then a com￾pensatory saccade which accurately brings gaze onto the remembered target location is generated (head-fixed sac￾cade32 and head-free gaze shifts33). Strikingly, such compen￾sation does not require visual or proprioceptive feedback, because it can occur in complete darkness and after surgical deafferentation of extraocular proprioception in monkeys34. These results have, thus, revealed the existence of a non-sen￾sory feedback loop that can account for the remarkable ac￾curacy of the saccadic system under normal conditions35,36. For arm movements, the existence of non-sensory feed￾back loops that make use of motor outflow was initially sug￾gested by behavioral studies that showed that hand trajec￾tory could be amended with a shorter latency than the minimal latency required to process peripheral information. For instance, Higgins and Angel37 observed that the reac￾tion time to an unexpected modification of the target trajectory in a manual tracking task was shorter than the proprioceptive reaction time. A similar result was reported by Jaeger et al.38, who found that altering the proprioceptive signal through vibration did not modify the reaction time to a visual perturbation. Using a pointing task, Cook and Diggles39 observed corrections to the hand path within 45 ms when the initial direction of movement was incor￾rect. This value was close to that reported by van Sonderen et al.40 (30 ms) in a double-step task in which the initial tar￾get location was changed during or after the initiation of movement. Convincing arguments were reported in an elegant study by Wolpert et al.29, which suggested that a reliable es￾timation of the location of the hand could be obtained by combining efferent and afferent signals in a forward model. These authors required subjects to move their hand along a line while holding a manipulator. The hand was allowed to be viewed for 2 s prior to the onset of movement. The ma￾nipulator was connected to a torque motor that induced re￾sistive or assistive force to the movement. At the end of the trial, the subjects estimated the location of their hand using a visual spot controlled by the other hand. The temporal propagation of measured errors exhibited by the subjects could be fully accounted for by assuming that the motor control system integrates both motor outflow and sensory inflow to estimate the location of the hand. By contrast, models based exclusively on either sensory inflow or motor Review Desmurget and Grafton – Forward modeling 426 Trends in Cognitive Sciences – Vol. 4, No. 11, November 2000 trends in Cognitive Sciences Desired displacement Intitial conditions Feedback module Inverse model Motor plan Final state estimate (end-point location) Forward model Motor outflow sensory inflow Continuous control ES Corrective command Motor command Hand location Target location Fig. 2. Simplified hybrid circuits that makes use of a forward model of arm dynamics for controlling hand movements. A motor plan is initially defined based on the respective locations of the hand and target. During the movement, a forward model of the dynamics of the arm is generated. This model receives the sensory inflow and a copy of the motor outflow as inputs and generates an es￾timate of the movement end-point location as output. This estimate is compared to the target location. In case of discrepancy, an error signal (ES) is generated, triggering a modulation of the ongoing motor command

De5网gr9t0n6年F44行gn一手r群4四gd号 Review ouflow were unble to predict the cheerved pattern of Arpimens that sppeet this dlaim can be found in sudie ertot.Hoe and Arbih teached a comeatible candluuion. howing that viion of the hand at re,prior o movemeat. Th灯ood,or reaching movements.thac conmol mo出 impeowes mevement acouracy through an eptimizarien of chat combine cfferent sigruls and aferent informatioe to e anline fecdback loopu“ cimare the cumenr lecarion of che hand and adjusr the The snudies reviewed in the foregoing secrion have es planed pem of masde activatio y captud blihed that forad modds can combine mote ouw che kinematic characteristics of visaally directed reaching and sensory intlw to esimale the cutrent and futue states In particular,thi model which ued a look-ahead predictor of the motor app with nepigibie dely.Thee fiad. o cempensare for delays,was able oo accounr for the mrafec. ings therefore.blae the kry argmenr the use ory reviion thut is oberved n behvioral cperimentsi which the tarprt lcati modifid at the beginaingof the hand movemeat or daring the saccadic reponse. Noa-qential,on-bllisic conl of reaching Purther argumensn free of che cendsion thar effer Twe ooncepes.squenti controf and ballisic ring ent and乐ereat aig山e emhined mgeeehe forward estimarion have been peevided by necent studies on are generated over time.Foe an exenaal ccaminer.the rela interjoiat coordinion.Gibble and Ctrytt showed that tive ccedinion of the eye.head and hand daring oal dectromyographic (EMG)viry in the souder and direced eching ppears to be soquei Whena bject cho joints vari in a peodictive mner duing reaching points to a vil targe in peripheral apacr,the eye move movemenes to compeasane for inceracion porque that arises first.followed by the head and finally the hand.The gaoear fom multijoint dynamics.Thi aneicipaory tepome indi rives at the target before ce at about the same time a the cares that the nervous srem can use a foewand moddl to orege of the hand movemere.hecause the duration of eve predict and ff the kinematic cmquences of iareg- mowement is hrid Severa researchers hypotheiaed that mena dynamcs.Inerestingly.recent results indicare that this soquential ergininron has a fnctionl foundcon eraory infoemarinn is cricical so uet parameren foe and up- and comoqueatly wrpeed that the acrvoen yuem had to dare sch a forward medd For cample.Saiabung achieveget foverion befoe beildnga relibie mo pln uird to paticns th presemedfiber smory for the arm.becae the extra-foveal viual signal did rot ncurpathy to make a geture similar to sliing a la of allw for an acurae timtion of the target loction bread.Whou being able mo se heir limb moiag these Thi hypecheis wa chilleaged by mi th shoued that purienes could noc compensace foe imeractien corque. the seril enganicadon of the cye.head and amm ar the beh which lod to evere impuirments in interjoint coordiration ioral kvel set primarily fom inertial facton.A shown by The best eridence for the presence of a forward modd Biguer wats,the EMG dischange for the ee.head and arm during eaching was prorided mce鸣b可Bard"A during fast reaching mowmenes is nearly synchmnous indi- deafferentod patiene was inserucod to lok at and point o cating thit the mooor command is sent to these ditferent ef vil arethat were diplayed in the peripheral ficld of fes in puralld (the arm moves lat smply becae it ha viin.butws not alked to kok at the movin limb.Ia the greatest inema球lf0 ae considers thar the0at国ng some triah,the tat locationa changed lightly daring aristic mude contraction occun 50-100 m befoee the ac. the coae of the ocuar saccade.Saccadic supprenion pre tual motio in a seaching movement",this obecrvation con. venred the patient from c detecting thia manipu. cun with peychopkyical i th he demonrad that aio whos coavinced that she peimod ry the movement generaly follkws the socdic respomse eat.The patieat wa able to crt her movement online Wih1ga60o1间wR五1346 o neach the new ranger lecarion despine the absence of pe The abe coned ebservarions indicare thar the ini ripheral infermarion.It is impeetant to noce tha her com mmand can be isued on the bais of an impee oction were not as accurate n thoe ol control subjects. fect etimtin of the targer loction A the end of the ocu which that moe ouflow hado be cmbined ih第le除ensory indow0a家noptim initiaed.the indal esdimatien of che ranget posirion is up inate of locatioe ol the hand. 由ied on the mis of foveal information.始p山ting is Several sdies have amempoed to nare the pame of dlearly demenrad by the fiading thar amm movemens end-poine with key varisbles that ued to pha ml36 rae when th0atw0mme山 sehing movement.For insunce.Vindras era eWe replicaed and eaended this fiding eny showed that firl ermon in villy-directed movemerta that (anpahlibed),uinga preocl ad apprm tothe were performed wichour being sble ro view the meving limb one described in Prablac Pocusing on hand kine. eeflcted wptematie biaes in the etimeion of the initi matic,we oberved the following.Fint,the initial acceke. state of the motor apparaes.Thi kind of rewlt might ap ation vectoe ws otadeted by the bility to moe the cyes pear to contradict the idea that moement trajectary is con- which cnncun with the idea that the initial moror plan is crolled by intemal feedback loops.However.this is noc the asembld on the basis of periphera visal infoemarion casc,becaae the extimutian of the current locatioe of the Second the masiml vdociry veetoe had a smaller mapi- hand by a forwand model will be amecred in a systemaric nude in the eye-feoe condirion than in the eye-ficed condi. manner if either the estimaion of the initia ware of the rion.which is in aperment with the observarion thar mobee apparanus6鲜the imerse modd that transfor th女 distance are usually overestimated in the peripheral viu desired diplcement into a moece command,is biaed fiedu,Thee obeerrations wppeut that an inaccurate 27 Trend in Cognitive Sciences Vol.4.Na.11.Navember 2000

outflow were unable to predict the observed pattern of error. Hoff and Arbib14 reached a compatible conclusion. They showed, for reaching movements, that control models that combine efferent signals and afferent information to es￾timate the current location of the hand and adjust the planned pattern of muscle activation successfully captured the kinematic characteristics of visually directed reaching. In particular, this model, which used a look-ahead predictor to compensate for delays, was able to account for the trajec￾tory revision that is observed in behavioral experiments in which the target location is modified at the beginning of the hand movement or during the saccadic response. Further arguments in favor of the conclusion that effer￾ent and afferent signals are combined to generate a reliable forward estimation have been provided by recent studies on interjoint coordination. Gribble and Ostry41 showed that electromyographic (EMG) activity in the shoulder and elbow joints varies in a predictive manner during reaching movements to compensate for interaction torque that arises from multijoint dynamics. This anticipatory response indi￾cates that the nervous system can use a forward model to predict and offset the kinematic consequences of interseg￾mental dynamics. Interestingly, recent results indicate that sensory information is critical to set parameters for and up￾date such a forward model. For example, Sainburg et al.20 required two patients that presented with large-fiber sensory neuropathy to make a gesture similar to slicing a loaf of bread. Without being able to see their limb moving, these patients could not compensate for interaction torque, which led to severe impairments in interjoint coordination. The best evidence for the presence of a forward model during reaching was provided recently by Bard et al.18 A deafferented patient was instructed to look at and point to visual targets that were displayed in the peripheral field of vision, but was not allowed to look at the moving limb. In some trials, the target location was changed slightly during the course of the ocular saccade. Saccadic suppression pre￾vented the patient from consciously detecting this manipu￾lation, who was convinced that she pointed to a stationary target. The patient was able to correct her movement online to reach the new target location despite the absence of pe￾ripheral information. It is important to note that her cor￾rections were not as accurate as those of control subjects, which suggests that motor outflow had to be combined with at least some sensory inflow to generate an optimal es￾timate of location of the hand. Several studies have attempted to relate the pattern of end-point errors with key variables that are used to plan reaching movements2,42,43. For instance, Vindras et al.44 showed that final errors in visually-directed movements that were performed without being able to view the moving limb reflected systematic biases in the estimation of the initial state of the motor apparatus. This kind of result might ap￾pear to contradict the idea that movement trajectory is con￾trolled by internal feedback loops. However, this is not the case, because the estimation of the current location of the hand by a forward model will be affected in a systematic manner if either the estimation of the initial state of the motor apparatus or the inverse model that transforms the desired displacement into a motor command, is biased. Arguments that support this claim can be found in studies showing that vision of the hand at rest, prior to movement, improves movement accuracy through an optimization of online feedback loops45. The studies reviewed in the foregoing section have es￾tablished that forward models can combine motor outflow and sensory inflow to estimate the current and future states of the motor apparatus with negligible delays. These find￾ings, therefore, obliterate the key argument against the use of feedback mechanisms for fast reaching movements. Non-sequential, non-ballistic control of reaching Two concepts, ‘sequential control’ and ‘ballistic reaching’, have strongly influenced our thinking of how motor plans are generated over time. For an external examiner, the rela￾tive coordination of the eyes, head and hand during goal￾directed reaching appears to be sequential. When a subject points to a visual target in peripheral space, the eyes move first, followed by the head and finally the hand. The gaze ar￾rives at the target before or at about the same time as the onset of the hand movement, because the duration of eye movement is brief 13,46. Several researchers hypothesized that this sequential organization has a functional foundation23,46 and consequently suggested that the nervous system had to achieve target foveation before building a reliable motor plan for the arm, because the extra-foveal visual signal did not allow for an accurate estimation of the target location46,47. This hypothesis was challenged by studies that showed that the serial organization of the eye, head and arm at the behav￾ioral level results primarily from inertial factors. As shown by Biguer et al.5,48, the EMG discharge for the eye, head and arm during fast reaching movements is nearly synchronous, indi￾cating that the motor command is sent to these different ef￾fectors in parallel (the arm moves last simply because it has the greatest inertia). If one considers that the onset of an ag￾onistic muscle contraction occurs 50–100 ms before the ac￾tual motion in a reaching movement49, this observation con￾curs with psychophysical studies that have demonstrated that the arm movement generally follows the saccadic response with a lag of 60 to 100 ms (Refs 13,46). The abovementioned observations indicate that the ini￾tial motor command can be issued on the basis of an imper￾fect estimation of the target location. At the end of the ocu￾lar saccade, after commands for arm movements have been initiated, the initial estimation of the target position is up￾dated on the basis of foveal information. This updating is clearly demonstrated by the finding that arm movements are less accurate when the eyes are not free to move to the target46,47. We replicated and extended this finding recently (unpublished), using a protocol and apparatus similar to the one described in Prablanc et al.46 Focusing on hand kine￾matics, we observed the following. First, the initial acceler￾ation vector was not affected by the ability to move the eyes, which concurs with the idea that the initial motor plan is assembled on the basis of peripheral visual information. Second, the maximal velocity vector had a smaller magni￾tude in the eye-free condition than in the eye-fixed condi￾tion, which is in agreement with the observation that distances are usually overestimated in the peripheral visual field46,47. These observations suggest that an inaccurate Desmurget and Grafton – Forward modeling 427 Trends in Cognitive Sciences – Vol. 4, No. 11, November 2000 Review

Review Desmur9t第4G1a158n-Frwa7d■odaling (a) second targetaun (FgTrajectory amendment wese Open loop poinong , 500 500 mooth越in适ated by the ahence of discmetinaities in 20*30 wrist veleciry curves,which ehbined the sme bell-shaped prolile for both the pertubed and cumel movemenes. 300 020+30 Intereingly,were desectahle upto10ma f the onset of hand movemen.which wis roughly symcho 100 nbed with the end of the ocular saccade,suggesring thar 100 had ywa amendod very carly in the mvement. 0100300 500 The panem ef comection and the reaction time to the per- 0100300500 Distance (mm) Distance (mm) turbione imlar,epecive of whedher the ming (c) (d) limb cnuld he viewed.Thi y modifi- 260 500 cation that were abened in the douhle-aep trias muainl 20+10 200 20 depended on mon-visal feechack loops,which is comparble with obecrvations mertioned carict that showed the viaiun 160 of the limh duing the fir haf of the trajecory ha no effoc 100 星富 o mevement final accurac However,the geneiry of 100 this condusion was challenged by Band and colleagges who 0 0 deigted a tcries of caperimne hnod on the idea that pe 0 4008001200 40000 1200 ripheral viso通mighr be pa山dy sensrive to the dinecrion Time (ms) Time (ms) of the mevemen In their fiet operiment,the authors w雨 movements (130 ma)could be coerecnod under the coatrol maeth and ear作sh6 ettons during reachin同med witheut of perpheral viien even whea vhien was only availbl dur s的时he动线Menn同h国。a闲中特w梦高&*4y ing the fine halfof the trajectory.In light of these and other 唐-dCe4 onding velori时snf8 ded with perrsiasion from Ref.4】 dara'.the of end-peint accuracy that is chserved when visual infermation is remeved at the beginning of a mooor plan aemed onthe bsis of an imperfecr esme movement might be explinod by小eaha国heoe of the taget lcatioe is currected catly in the coune of the command can still be optimined by poeent erminal feedhack mevemenr.specifically during the accelerarien phase.This cacion dearly challenges the idea that vially directed mowemenrs ane haliwic dueing their carly stage. Feedfarwaed specification of the motor command To isverigare diretly whether an ineended action is a Given the powerful ability of forward models to adjust result of a preser pamem of nonmodifiable commands movemear oaline.ene might wonder whether any of the Prablane and colleapue doigned a se of behvioral caperi- movement necds to be plannod in advance.Foe cuam. mears in which the inirial inaccurae esimacion of the rangrt ple,it has heen shoan char a contol scheme thar iroles a location dering movement plarning anificially in. progressive definition of the motor cmndwidout creased.To achieve this.the authors sed a doublessrep poining paradigm in which the tarpt eiona modificd ture the kisema charaeeritics o viall-diseesed reach- dighly (tag jumpparadigm)during the cous of the ingg".This compuranonal result.however.is nor echood by oculat saccade.This prucedure hn the folkwing theee majee mtal findings.Two main lines of evidence advanragee (1)the tangrt jamp is noe perceied conscioudy 可the whject,becaue of ccadic山ppei的the tangs jump does国rer the organi20nofe0 culmoro号y The fest line of evkdence comes froe the soudy of fine tem,becaue aaccadic tpamc to worury targets imohe predictive commpematory adjutments in single mudles. an inirial sccade thar uadenhoons the tanget positien and a Gebble and Oary"roady repod that EMG acriviry in singl cotrective sccaden to achicve accutate target aoquii the shoulder and dbow joints vriod in a peodictive manner tioe:and (the target jemp dees not alter the oeganzation to comtpesate for intcraction torqee that aroes foom muki o the manl pepomae,hecaine pointing movementx to jeint dymamic Sinib were repeeted for thedu rionary rargersenpare of the targer lcatin the ment of gip foe during arm movemens perommed with end of the ccadic diplacemene.which a taken ino acrour a hand-hed load and for pouural compention t to amend the ompoi军arm mowemeng (ee abo小，Thw hliae rapid amm movemer,Sach adjumens can anly poin can be summid by sing that uncomcodou be explained if the kinemcic comoquences of the upoom bleandnare identica fmfuncrion ing moor command can be peodicd peecisdy.th poim of view.The intraaccadc modification of the target motur command ia,to aome exeat,known'in advance. lecarien simply increases an emor thar is aleady present in The second ine of evidence that eveals the peeplan the tem which spors the hypothen at the same co ning peoccss come from a tocent trameranial magnctic reciw proces ane engagnd in the 'jumpand'stationary' rimuarion (TMS-neachin军动that was carrind ou可 trale Uing the tarpet jump paradip,Prabane and c our poup2”g4线.Siec率idwith leges oerved that the hand path which was initiy di their right hand bur vision of the arm wis noc allod dar. red the fint tarpet.dived poyd the ing this movement.In some trial,the target locarion wa Trends in Cognitiee Stlantes Yal.4.Ne.11

motor plan assembled on the basis of an imperfect estimate of the target location is corrected early in the course of the movement, specifically during the acceleration phase. This conclusion clearly challenges the idea that visually directed movements are ballistic during their early stage. To investigate directly whether an intended action is a result of a preset pattern of non-modifiable commands, Prablanc and colleagues designed a set of behavioral experi￾ments in which the initial inaccurate estimation of the target location during movement planning was artificially in￾creased12,13. To achieve this, the authors used a double-step pointing paradigm in which the target location was modified slightly (target ‘jump’ paradigm) during the course of the ocular saccade. This procedure has the following three major advantages: (1) the target jump is not perceived consciously by the subject, because of saccadic suppression; (2) the target jump does not alter the organization of the oculomotor sys￾tem, because saccadic responses to stationary targets involve an initial saccade that undershoots the target position and a single corrective saccade35 to achieve accurate target acquisi￾tion; and (3) the target jump does not alter the organization of the manual response, because pointing movements to sta￾tionary targets involve an update of the target location at the end of the saccadic displacement, which is taken into account to amend the ongoing arm movement (see above). These points can be summarized by saying that unconscious dou￾ble- and single-step situations are identical from a functional point of view. The intrasaccadic modification of the target location simply increases an error that is already present in the system, which supports the hypothesis that the same cor￾rective processes are engaged in the ‘jump’ and ‘stationary’ trials. Using the target jump paradigm, Prablanc and col￾leagues observed that the hand path, which was initially di￾rected to the first target, diverged progressively toward the second target12,13,15 (Fig. 3). Trajectory amendments were smooth, as indicated by the absence of discontinuities in wrist velocity curves, which exhibited the same bell-shaped profile for both the perturbed and control movements. Interestingly, corrections were detectable up to 110 ms after the onset of hand movement, which was roughly synchro￾nized with the end of the ocular saccade, suggesting that hand trajectory was amended very early in the movement. The pattern of correction and the reaction time to the per￾turbation were similar, irrespective of whether the moving limb could be viewed13. This suggests that trajectory modifi￾cations that were observed in the double-step trials mainly depended on non-visual feedback loops, which is compatible with observations mentioned earlier that showed that vision of the limb during the first half of the trajectory has no effect on movement final accuracy24,25. However, the generality of this conclusion was challenged by Bard and colleagues, who designed a series of experiments based on the idea that pe￾ripheral vision might be particularly sensitive to the direction of the movement6,23. In their first experiment, the authors showed that the directional component of very fast aiming movements (,130 ms) could be corrected under the control of peripheral vision even when vision was only available dur￾ing the first half of the trajectory50. In light of these and other data6 , the preservation of end-point accuracy that is observed when visual information is removed at the beginning of a movement might be explained by the fact that the motor command can still be optimized by potent terminal feedback loops. Feedforward specification of the motor command Given the powerful ability of forward models to adjust movement online, one might wonder whether any of the movement needs to be planned in advance10,11. For exam￾ple, it has been shown that a control scheme that involves a progressive definition of the arm motor command, without any preplanning adequately predicts trajectories that cap￾ture the kinematic characteristics of visually-directed reach￾ing11. This computational result, however, is not echoed by other experimental findings. Two main lines of evidence suggest that a representation of the upcoming motor com￾mand exists prior to the onset of reaching movements. The first line of evidence comes from the study of fine predictive compensatory adjustments in single muscles. Gribble and Ostry41 recently reported that EMG activity in the shoulder and elbow joints varied in a predictive manner to compensate for interaction torque that arises from multi￾joint dynamics. Similar results were reported for the adjust￾ment of grip force during arm movements performed with a hand-held load51 and for postural compensations that sta￾bilize rapid arm movements52. Such adjustments can only be explained if the kinematic consequences of the upcom￾ing motor command can be predicted precisely, i.e. if this motor command is, to some extent, ‘known’ in advance. The second line of evidence that reveals the preplan￾ning process comes from a recent transcranial magnetic stimulation (TMS)-reaching study that was carried out by our group15 (Fig. 4). Subjects pointed to visual targets with their right hand, but vision of the arm was not allowed dur￾ing this movement. In some trials, the target location was Review Desmurget and Grafton – Forward modeling 428 Trends in Cognitive Sciences – Vol. 4, No. 11, November 2000 trends in Cognitive Sciences Open loop pointing 500 300 100 0 0 100 300 500 10 20 20 30 250 150 200 100 50 0 0 400 800 1200 Planar hand path (mm) Hand tangential velocity (deg. sec–1) Distance (mm) Time (ms) 20 30 500 300 100 0 0 100 300 500 Planar hand path (mm) Distance (mm) 20 10 500 300 100 0 0 400 800 1200 Hand tangential velocity (deg. sec–1) Time (ms) (a) (b) (c) (d) Fig. 3. Smooth and early path corrections during reaching performed without vision of the limb. (a) Mean hand path of a subject reaching towards a stationary target (208; broken line) and displaced targets (208 to 308 and 208 to 108; unbroken lines). (b–d) Corresponding velocity profiles. (Figure reproduced, with permission, from Ref. 74.)

Desmurget and Gratton Forward modeliag Review displaced during the socadic respore.wheres in other tri No-stimulation semned stationry.As observeder Let stimulalion nght hand right hand the taret jump dicined a smh and peopenive adju ment of the hand path.Smrikingly.when a sngle TAiS was pplicd over the lft imraparicd sukus (IPS)at the onct of hand movement,these amooth path coerections wene di npted and the mbjoct poireed o the fot aget locatinn However.the hand majeeyos did not become craic,pevag ativey ccumave nn移can be perf而med in the abornce of online feedhack loops If conin cor lo,which the rdlative costhe hand and targer ae cmd wered enerare the motoe command in teal time,dirupting thee oops shoud he reshed in ther errant or dramatically F与4 Critieadl cele of the pesterbor4etad6earr神e4g0 inaccrtsry targe.However,acon 8 ean hand pathe of one subject,aither wich0tre08时hout (middie pan回 eml seuion that imolved the tame stimulrion site bat the transtranial mugnetic stireu山ses.Be年ontinuou mu时peesent the mean ga other hand (lf hand,ipirerl o the stimalated aite dinectad at jumping tang电好e22学and1goT】.ad0辆ndcate ctationary ta faled reveal any dsption of hand pach comrecrions This indicatea that the feedback dntupeion obecrvod for the 转anelL delerrsired try theee-dmerolenl MR0t时.的mtn5s年d凤 righe hand was not related to oculemeeer deficins or to the path com过 diextedd at satiomary tatget beoame less aoourate [aithonagh not errati inablity of the subject to update the target loction. Unfory the acoy and the degree ofdetail of the isitial motar plan cannot be etabished from the ahove. The sexls indicated that the pacicnt wax ahle to reach the mentioned dara.The fact thar the moror defidirs observed in ager properly in the saony cendicion.bur she pre ome dealferened paticnts cannot be caplained solely by sented a dramatie inability to coerect her ongoing mowe. dheir incapaciy m dfine the inirial ane of the motr me in the perurbed cnditin.In thercthe pa that the initial plan is only cruddly tient pointed pmnerally to the inicial t cn befare defined prioe oo the ensct of movement and subsoqueady initiating a socond movement to the finall targer posirion up山ted through inema foedback bop重he movement Carerl subjeca ethbited cacly modficatios of the hand progress.Theloss of accuracy thar isobd when caline path,as epeced from aier eedback学are dsnted sppons this views(Fig绿 Recent studio in the meakry and human have shown that the parictall corex is highly diferentied with miny Functional anmy ofiaerae fanctinal nhdviion Unfoetunay,the fusction of Movemenes sch as eching are controed by wdely dis many of these subdivisions is nec earirely undersoood and is ributed cortical and subcortical acmceimoor aren.It is noe dlea which pecific areas might be auociated with imer. gmerally hough that parieral and prem are nal feedbad leps In addirion.peogress in this anea is ham foe the ecion pparinn and i of a pend by the marked differences betwnen the paricral be movement.Although the functienal anatomy of inoerna cymoaciteconi orginirion of haman and non-umun fedback loop n fally known,two areas withn the dis primareDapine thee chtace wo potenti f the tribured seasormecor system are hypothesaed to be critical PPC hve been evoked in the literane,paricularly in areas for updating hand traectory,samely the PPC in the regioe that imole the IPS.It wa fnt wed th the PPC might of the IPS and the anterior paraniggintal coetes of the cet. beimodincmutingbyminthec cbelum挑，In thisc吗min the poeential tual location wich the o the handn.This view criburion of chese rwo srucnenes to inemal feocback. is suppered mainly by the demerstraron thar the PPCC i Indirect evidence to uppevt that isternal feedhack crucial for athihing stahie telaticships beween hetero- oops rel中en the PPC comes from the中ervation thar ch geneo inomo for mergng the ar and agt. ech teipemiveness of ncurons in area 7a of non-hum related ignals into a commom frame of refeence,which con. primates changes as the umeen hand approaches a vhul stioates a crucal sep in defaing a mecor erree.An aliemudive apr“A more convining arpmeat is proided b可c (but not echive)bypeeheis peoponcs that the PPC pene TMS sdy described in the previous sectien.When the ares a forwand model of the locarion of the hand.Indirect omm fanctioning o the PPC in the repion of the IPS is wapport for thi vicw come from the olervation that several perurbed after the omet of hand movemere.feechack loops that allow cnmrection of the ongoing movement are g山nia,5 enngr appor emerped wihd女c disrupted(Fg 4).This focall deficir was recencly epli inding that sensorysgak fomm呼nodaese长iod cated in dinical atudies that irolved a patient that pre- p军cptive,audary and veeb山r以well ax cfferent senoed wich biareral ischemic kesions of the PPC (Ref.57) copy sigrals from moor suces are inegrad in hePPC This patient was asod to look at and point to visual targets (Ref.6)Thi coecun with the idea that scmoori-motor inte presented on a computer screen in front of her.In some tri. on isacraci fere of foeward modch. lthe taget remaned ry wherea in other tral it rioned findin零that impicae the jumped'to a sew lecarion ar the oasec of arm movemear. PPCin feedforward modek.seveal linesofevidence suppon Tre4t1sCee角itive54e44年Val,4,Ne。11。N0 vember Ja00

displaced during the saccadic response, whereas in other tri￾als, it remained stationary. As observed in earlier studies12,13, the target jump elicited a smooth and progressive adjust￾ment of the hand path. Strikingly, when a single TMS was applied over the left intraparietal sulcus (IPS) at the onset of hand movement, these smooth path corrections were dis￾rupted and the subject pointed to the first target location. However, the hand trajectory to stationary targets did not become erratic, suggesting that relatively accurate move￾ments can be performed in the absence of online feedback loops. If continuous control loops, in which the relative lo￾cations of the hand and target are compared, were used to generate the motor command in real time, disrupting these loops should have resulted in either errant or dramatically inaccurate trajectories to stationary targets. However, a con￾trol session that involved the same stimulation site but the other hand (left hand, ipsilateral to the stimulated site) failed to reveal any disruption of hand path corrections. This indicates that the feedback disruption observed for the right hand was not related to oculomotor deficits or to the inability of the subject to update the target location. Unfortunately, the accuracy and the degree of detail of the initial motor plan cannot be established from the above￾mentioned data. The fact that the motor deficits observed in some deafferented patients cannot be explained solely by their incapacity to define the initial state of the motor sys￾tem2,20,53 suggests that the initial motor plan is only crudely defined prior to the onset of movement and subsequently updated through internal feedback loops as the movement progresses. The loss of accuracy that is observed when online feedback loops are disrupted supports this view15 (Fig. 4). Functional anatomy of internal error corrections Movements such as reaching are controlled by widely dis￾tributed cortical and subcortical sensorimotor areas. It is generally thought that parietal and pre-motor systems are essential for the selection, preparation and execution of a movement. Although the functional anatomy of internal feedback loops is not fully known, two areas within the dis￾tributed sensorimotor system are hypothesized to be critical for updating hand trajectory, namely the PPC in the region of the IPS and the anterior parasagittal cortex of the cer￾ebellum54,55. In this section, we examine the potential con￾tribution of these two structures to internal feedback. Indirect evidence to suggest that internal feedback loops rely on the PPC comes from the observation that the reach responsiveness of neurons in area 7a of non-human primates changes as the unseen hand approaches a visual target56. A more convincing argument is provided by the TMS study described in the previous section. When the normal functioning of the PPC in the region of the IPS is perturbed after the onset of hand movement, feedback loops that allow correction of the ongoing movement are disrupted15 (Fig. 4). This focal deficit was recently repli￾cated in clinical studies that involved a patient that pre￾sented with bilateral ischemic lesions of the PPC (Ref. 57). This patient was asked to look at and point to visual targets presented on a computer screen in front of her. In some tri￾als, the target remained stationary, whereas in other trials it ‘jumped’ to a new location at the onset of arm movement. The results indicated that the patient was able to reach the target properly in the stationary condition, but she pre￾sented a dramatic inability to correct her ongoing move￾ments in the perturbed condition. In the latter case, the pa￾tient pointed generally to the initial target location before initiating a second movement to the final target position. Control subjects exhibited early modifications of the hand path, as expected from earlier studies12,13,15. Recent studies in the monkey and human have shown that the parietal cortex is highly differentiated with many functional subdivisions58–60. Unfortunately, the function of many of these subdivisions is not entirely understood and it is not clear which specific areas might be associated with inter￾nal feedback loops. In addition, progress in this area is ham￾pered by the marked differences between the parietal lobe cytoarchitectonic organization of human and non-human primates59. Despite these obstacles, two potential roles for the PPC have been evoked in the literature, particularly in areas that involve the IPS. It was first suggested that the PPC might be involved in computing motor errors by comparing the ac￾tual target location with the location of the hand15. This view is supported mainly by the demonstration that the PPC is crucial for establishing stable relationships between hetero￾geneous information58, i.e. for merging the arm- and target￾related signals into a common frame of reference, which con￾stitutes a crucial step in defining a motor error. An alternative (but not exclusive) hypothesis proposes that the PPC gener￾ates a forward model of the location of the hand. Indirect support for this view comes from the observation that several parietal structures are concerned with various types of predic￾tive mechanisms61–63. Stronger support emerged with the finding that sensory signals from many modalities (e.g. visual, proprioceptive, auditory and vestibular), as well as efferent copy signals from motor structures, are integrated in the PPC (Ref. 64). This concurs with the idea that sensori-motor inte￾gration is a crucial feature of forward models. Despite the abovementioned findings that implicate the PPC in feedforward models, several lines of evidence support Desmurget and Grafton – Forward modeling 429 Trends in Cognitive Sciences – Vol. 4, No. 11, November 2000 Review trends in Cognitive Sciences Left stimulation right hand No-stimulation right hand Fig. 4. Critical role of the posterior parietal cortex for online movement correc￾tions. Mean hand paths of one subject, either with (left panel) or without (middle panel) transcranial magnetic stimulation (TMS). The continuous curves represent the mean paths directed at stationary targets (208, 308 and 408). The dashed curves represent the mean paths directed at jumping targets (308 to 22.58 and 308 to 37.58). Black circles indicate stationary tar￾get locations and white circles represent jumping target locations. Location of TMS site (right panel), determined by three-dimensional MRI (black circle). When TMS is applied, path corrections that normally occur in response to the target jump are disrupted. In addi￾tion, movements directed at stationary targets become less accurate (although not erratic). (Figure reproduced, with permission, from Ref. 15.)

Review Desmurget and Grafton Forwaed modeling the idea thar forwrards models mighe aely on the cerebellam natue of their mciprocal interacticns and their poeenrial more thas the PPC.Urlike the PPC,the cerebelum ha coanections with ocher areas,remain to be inveiganed. long beendh feodbacko Theoriginal idea,propoed b句Hdma“,aththiatrucisetic Condluding remarks ulrly imporant in the vi pidance ofmm This The presear review indicanes that no single compurarional hppothesn was socently i送porated into a mo家cl algoeichm can adoguately describe the coateol proceses tht cheme in which it issmod tht the cebelum plays are used to perform godirected movements.Rather. crucil role in dahorating forward and imene interaal reaching souard a target requne an innegratie contrl modds.Based on this idea (which has been revlewed cle cheme in which f6 dforward specficatieno国the mot国 where)oncdi two fuctios to the cerbelam commad,furward modelng of the dynamics of the arm for mewemear guidance.First.this struc might con and online updating of che inirial patern of munde acti trbuse to the comenion of the emor signal penerated by the vation are yheiaed in relible frdback lops whichan PPC into a mooor command (inverse crans-foemarioa).In thought to imolve the cerebellum and PPC. upport ofi thi view,it ha been hown that imene modds are endlosed within the cerebellumand thac pa ticnts with cerebdlat lesions are mable to comperoue lo ed击eheee drda年G1se向dde每d日B时ear线e mutijoint inerctdimovmenling Ad poeential functios of the cem is ioe ment in esmating the probable posirion and velociry of the cal aukctanoe.年pomed b句4whne cffoctor (farward model.Of the aupportive evidence. reviewed dehethe follwing three ideas are ee cialy comiacing Firt in puientsleion Beferenoes path comocions that ane bused on visua scroery indormation are characteriad by ccie deriatinn and aboemal oci lations s would be espected ffeedback loops ed edu 2BytM时图em eye to hand4yte时 ively on derpod semory inmmtion.Secend an noou ae)forward modd kads to tracking deficis OMutr Comoal fod.J1 that ae milar mo those oberved in cebelle araxia Finally.cerebellar parhelegy leads ro major deficits in motor ath中on the gencracion of a foeward modc.Fe四 cool.m eM长#nodn网 esample.when normal subjects hfr an chjecr.che load on the ,d华p1婚酸 hand increaes with mewement accekration and variacion in iration-af Coak p foece anticipae恤ncrene of load force为T色can orly he eplained if the kinematic comoqumces c an upcoming mecor command can be peodictod peecisely 4*74417 throa forward modd Imeretingly.the doe coupling beeweenp and load forces is bent from patienes with d4d时 cerebellar lesion 重， The abevemencioned data sueest thar incem feed. hack loop rely os boh the cerchellum ad the PPC.The M号eom a1- esacr functional ole of these rue stucnres,as well as the Outstanding questions 1MP01-OYTUQ .Classkal models of motor control divide reaching movements into two 12.D.etal components mamely a pre-programmed (or ballstic)component,which ensures rapid transport of the hand to the vicinity of the target.and a 神4y controlled component,which allows fine spatial adjuntments at the end of the trajectory.Is thi view still tenable? 14 中w d1月⊙4如PN coetrol the trajectory of the hand during teaching movements? +Forward models allew prediction af the behavior of a contrelled sytem (e.q.arm)in tesporde to a given cemmand in advance.What is the p4kt中a联置15-般 evidence that the nervous system uses this type of forward model to 55品时l端e山时中8gan control reaching movements? .What i the evidence that a motor plan is effectvely assembled prior to 6可，3W.6喝o the onset of movement?ls it plausible that the motor command is not C700 planned in advance,but rather ge erated in real time va intemal feedback loops? 0 1T1-. ns风《时ww4w4 m and peining with vhaael d feedhack loops that control ongoing movements? htp perhelutm .Are forwvard modeh ued in other cognitive domains? uftereated man Brain 1o线Sa 430 Trends in Cognitiee Stlences Val.4,No.11,Nowember 2800

the idea that forwards models might rely on the cerebellum more than the PPC. Unlike the PPC, the cerebellum has long been associated with feedback control28,65. The original idea, proposed by Holmes66, was that this structure is partic￾ularly important in the visual guidance of movement. This hypothesis was recently incorporated into a more general scheme in which it is assumed that the cerebellum plays a crucial role in elaborating forward and inverse internal models. Based on this idea (which has been reviewed else￾where30,31), one could assign two functions to the cerebellum for movement guidance. First, this structure might con￾tribute to the conversion of the error signal generated by the PPC into a motor command (inverse trans-formation). In support of with this view, it has been shown that inverse models are enclosed within the cerebellum30,67,68 and that pa￾tients with cerebellar lesions are unable to compensate for multijoint interaction torques during movement planning69. A second potential function of the cer-ebellum is involve￾ment in estimating the probable position and velocity of the effector (forward model). Of the supportive evidence, reviewed elsewhere28,30,70, the following three ideas are espe￾cially convincing. First, in patients with cerebellar lesions, path corrections that are based on visual sensory information are characterized by excessive deviations and abnormal oscil￾lations, as would be expected if feedback loops relied exclu￾sively on delayed sensory information71. Second, an erro￾neous (or absent) forward model leads to tracking deficits that are similar to those observed in cerebellar ataxia28. Finally, cerebellar pathology leads to major deficits in motor tasks that rely on the generation of a forward model. For example, when normal subjects lift an object, the load on the hand increases with movement acceleration and variations in grip force anticipate this increase of load force51,72. This can only be explained if the kinematic consequences of an upcoming motor command can be predicted precisely through a forward model. Interestingly, the close coupling between grip and load forces is absent from patients with cerebellar lesions73. The abovementioned data suggest that internal feed￾back loops rely on both the cerebellum and the PPC. The exact functional role of these two structures, as well as the nature of their reciprocal interactions and their potential connections with other areas, remain to be investigated. Concluding remarks The present review indicates that no single computational algorithm can adequately describe the control processes that are used to perform goal-directed movements. Rather, reaching towards a target requires an integrative control scheme in which feedforward specification of the motor command, forward modeling of the dynamics of the arm and online updating of the initial pattern of muscle acti￾vation are synthesized in reliable feedback loops, which are thought to involve the cerebellum and PPC. Acknowledgements We thank Helene Gréa, Dr Claude Prablanc and Dr Denis Pélisson for their helpful remarks and their critical reading of this manuscript. We are also grateful to Laura Payne for editing this manuscript and to Serge Terrones for technical assistance. Supported by Public Health Service Grants NS3504 and NS37470. References 1 Woodworth, R.S. (1899) The accuracy of voluntary movement. Psychol. Rev. Monogr. 3 (Issue 13) 2 Desmurget, M. et al. (1998) From eye to hand: planning goal-directed movements. Neurosci. Biobehav. Rev. 22, 761–788 3 Keele, S.W. (1981) Behavioral analysis of movement. In Handbook of Physiology, Section 1: The Nervous System (Motor Control) (Vol. 2) (Brooks, W.B., ed.), pp. 1391–1414, Williams and Wilkins 4 Arbib, M.A. (1981) Perceptual structures and distributed motor control. In Handbook of Physiology, Section 1: The Nervous System (Motor Control) (Vol. 2) (Brooks, W.B., ed.), pp. 1449–1480, Williams and Wilkins 5 Jeannerod, M. (1988) The Neural and Behavioral Organization of Goal￾Directed Movements, Clarendon Press 6 Paillard, J. (1996) Fast and slow feedback loops for the visual correction of spatial errors in a pointing task: a re-appraisal. Can. J. Physiol. Pharmacol. 74, 401–417 7 Meyer, D.E. et al. (1988) Optimality in human motor performance: ideal control of rapid aimed movements. Psychol. Rev. 95, 340–370 8 Milner, T.E. (1992) A model for the generation of movements requiring endpoint precision. Neurosci. 49, 487–496 9 Plamondon, R. and Alimi, A.M. (1997) Speed/accuracy trade-offs in target-directed movements. Behav. Brain Sci. 20, 1–21 10 Hinton, G. (1984) Parallel computations for controlling an arm. J. Mot. Behav. 16, 171–194 11 Flanagan, J.R. et al. (1993) Control of trajectory modifications in target-directed reaching. J. Mot. Behav. 25, 140–152 12 Pélisson, D. et al. (1986) Visual control of reaching movements without vision of the limb. II. Evidence of fast unconscious processes correcting the trajectory of the hand to the final position of a double step stimulus. Exp. Brain Res. 62, 303–311 13 Prablanc, C. and Martin, O. (1992) Automatic control during hand reaching at undetected two-dimensional target displacements. J. Neurophysiol. 67, 455–469 14 Hoff, B. and Arbib, M.A. (1993) Models of trajectory formation and temporal interaction of reach and grasp. J. Mot. Behav. 25, 175–192 15 Desmurget, M. et al. (1999) Role of the posterior parietal cortex in updating reaching movements to a visual target. Nat. Neurosci. 2, 563–567 16 Keele, S.W. (1968) Movement control in skilled motor performance. Psychol. Bull. 70, 387–403 17 Taub, E. et al. (1975) Deafferentation in monkeys: pointing at a target without visual feedback. Exp. Neurol. 46, 178–186 18 Bard, C. et al. (1999) Deafferentation and pointing with visual double￾step perturbations. Exp. Brain Res. 125, 410–416 19 Rothwell, J.C. et al. (1982) Manual motor performance in a deafferented man. Brain 105, 515–542 Review Desmurget and Grafton – Forward modeling 430 Trends in Cognitive Sciences – Vol. 4, No. 11, November 2000 Outstanding questions • Classical models of motor control divide reaching movements into two components, namely a pre-programmed (or ballistic) component, which ensures rapid transport of the hand to the vicinity of the target, and a controlled component, which allows fine spatial adjustments at the end of the trajectory. Is this view still tenable? • Does a delay in sensorimotor loops prevent the use of feedback to control the trajectory of the hand during reaching movements? • Forward models allow prediction of the behavior of a controlled system (e.g. arm) in response to a given command in advance. What is the evidence that the nervous system uses this type of forward model to control reaching movements? • What is the evidence that a motor plan is effectively assembled prior to the onset of movement? Is it plausible that the motor command is not planned in advance, but rather generated in real time via internal feedback loops? • What are the anatomical structures that are associated with internal feedback loops that control ongoing movements? • Are forward models used in other cognitive domains?

Desmurget and Grafton-Forward modeling Review v.n 512524 ctional contr .1519-158 n.C t aL (1991)Prope Haht.E. nd Mi 92)Prope 2000)A n D.et aL (19 20.350-357 and D( (197 82231023 642-509 .B.L et 3t.(19 and Westling.G (1984)Roles of glabro 0ta19700p frain 56.7-84 e1-5C Trends in Cagnitive Sciences-Vol.4.No.11.November 200 ③

20 Sainburg, R.L. et al. (1993) Loss of proprioception produces deficits in interjoint coordination. J. Neurophysiol. 70, 2136–2147 21 Hollerbach, J.M. (1982) Computers, brains and the control of movement. Trends Neurosci. 5, 189–192 22 Gerdes, V.G. and Happee, R. (1994) The use of internal representation in fast gold-directed movements: a modeling approach. Biol. Cybern. 70, 513–524 23 Paillard, J. (1982) The contribution of peripheral and central vision to visually guided reaching. In Analysis of Visual Behavior (Ingle,D.J. et al., eds), pp. 367–382, MIT Press 24 Beaubaton, D. and Hay, L. (1986) Contribution of visual information to feedforward and feedback processes in rapid pointing movements. Hum. Mov. Sci. 5, 19–34 25 Carlton, L.G. (1981) Processing visual feedback information for movement control. J. Exp. Psychol. 7, 1019–1030 26 Redon, C. et al. (1991) Proprioceptive control of goal directed movements in man studied by means of vibratory muscle tendon stimulation. J. Mot. Behav. 23, 101–108 27 Von Holst, E. and Mittelstaedt, H. (1950) Das reafferenzprinzip. Wechelwirkung Zwischen Zentralnerven system und peripherie. Naturwis. 37, 464–476 28 Miall, R.C. et al. (1993) Is the cerebellum a Smith predictor? J. Mot. Behav. 25, 203–216 29 Wolpert, D.M. et al. (1995) An internal model for sensorimotor integration. Science 269, 1880–1882 30 Wolpert, D.M. et al. (1998) Internal models in the cerebellum. Trends Cognit. Sci. 2, 338–347 31 Kawato, M. (1999) Internal models for motor control and trajectory planning. Curr. Opin. Neurobiol. 9, 718–727 32 Keller, E.L. et al. (1996) Endpoint accuracy in saccades interrupted by stimulation in the omnipause region in monkey. Vis. Neurosci. 13, 1059–1067 33 Pelisson, D. et al. (1995) Online compensation of gaze shifts perturbed by micro-stimulation of the superior colliculus in the cat with unrestrained head. Exp. Brain Res. 106, 196–204 34 Guthrie, B.L. et al. (1983) Corollary discharge provides accurate eye position information to the oculomotor system. Science 221, 1193–1195 35 Robinson, D. (1975) Oculomotor control signals. In Basic Mechanisms of Ocular Motility and their Clinical Implications (Lennerstrand, G. and Bach-y-Rita, P., eds), pp. 337–344, Pergamon Press 36 Guitton, D. (1992) Control of eye–head coordination during orienting gaze shifts. Trends Neurosci. 15, 174–179 37 Higgins, J.R. and Angel, R.W. (1970) Correction of tracking errors without sensory feedback. J. Exp. Psychol. 84, 412–416 38 Jaeger, R.J. et al. (1979) Directional errors of movement and their correction in a discrete tracking task. J. Mot. Behav. 11, 123–133 39 Cooke, J.D. and Diggles, V.A. (1984) Rapid error correction during human arm movements: evidence for central monitoring. J. Mot. Behav. 16, 348–363 40 Van Sonderen, J.F. et al. (1989) Motor programs for goal-directed movements are continuously adjusted according to changes in target location. Exp. Brain Res. 78, 139–146 41 Gribble, P.L. and Ostry, D.J. (1999) Compensation for interaction torques during single- and multijoint limb movement. J. Neurophysiol. 82, 2310–2326 42 Flanders, M. et al. (1992) Early stages in sensorimotor transformations. Behav. Brain Sci. 15, 309–362 43 Gordon, J. et al. (1994) Accuracy of planar reaching movements. 1. Independence of direction and extent variability. Exp. Brain Res. 99, 97–111 44 Vindras, P. et al. (1998) Pointing errors reflect biases in the perception of the initial hand position. J. Neurophysiol. 79, 3290–3294 45 Desmurget, M. et al. (1995) Representation of hand position prior to movement and motor variability. Can. J. Physiol. Pharmacol. 73, 262–272 46 Prablanc, C. et al. (1979) Optimal response of eye and hand motor system in pointing at visual target. I. Spatio-temporal characteristics of eye and hand movements and their relationships when varying the amount of visual information. Biol. Cybern. 35, 113–124 47 Bock, O. (1993) Localization of objects in the peripheral visual field. Behav. Brain Res. 56, 77–84 48 Biguer, B. et al. (1982) The coordination of eye, head and arm movements during reaching at a single visual target. Exp. Brain Res. 46, 301–304 49 Turner, R.S. et al. (1995) Directional variation of spatial and temporal characteristics of limb movements made by monkeys in a two￾dimensional work space. J. Neurophysiol. 74, 684–697 50 Bard, C. et al. (1985) Role of peripheral vision in the directional control of rapid aiming movements. Can. J. Psychol. 39, 151–161 51 Flanagan, J.R. and Wing, A.M. (1997) The role of internal models in motion planning and control: evidence from grip force adjustments during movements of hand-held loads. J. Neurosci. 17, 1519–1528 52 De Wolf, S. et al. (1998) Anticipatory postural adjustments during self-paced and reaction-time movements. Exp. Brain Res. 121, 7–19 53 Ghez, C. et al. (1995) Impairments of reaching movements in patients without proprioception. II. Effects of visual information on accuracy. J. Neurophysiol. 73, 361–372 54 Desmurget, M. et al. (1999) Functional anatomy of error correction during reaching movements in the dark. Soc. Neurosci. Abstr. 25, 787 55 Inoue, K. et al. (1998) PET study of pointing with visual feedback of moving hands. J. Neurophysiol. 79, 117–125 56 MacKay, W.A. (1992) Properties of reach-related neuronal activity in cortical area 7A. J. Neurophysiol. 67, 1335–1345 57 Pisella, L. et al. (2000) An ‘automatic pilot’ for the hand in human posterior parietal cortex: toward reinterpreting optic ataxia. Nat. Neurosci. 3, 729–736 58 Colby, C.L. (1998) Action-oriented spatial reference frames in cortex. Neuron 20, 15–24 59 Milner, A.D. and Dijkerman, H.C. (1998) Visual processing in the primate parietal lobe. In Comparative Neuropsychology (Milner, A.D., ed.), pp. 70–94, Oxford University Press 60 Sakata, H. et al. (1997) The TINS Lecture. The parietal association cortex in depth perception and visual control of hand action. Trends Neurosci. 20, 350–357 61 Eskandar, E.N. and Assad, J.A. (1999) Dissociation of visual, motor and predictive signals in parietal cortex during visual guidance. Nat. Neurosci. 2, 88–93 62 Duhamel, J.R. et al. (1992) The updating of the representation of visual space in parietal cortex by intended eye movements. Science 255, 90–92 63 Colby, C.L. et al. (1996) Visual, presaccadic, and cognitive activation of single neurons in monkey lateral intraparietal area. J. Neurophysiol. 76, 2841–2852 64 Andersen, R.A. et al. (1997) Multimodal representation of space in the posterior parietal cortex and its use in planning movements. Annu. Rev. Neurosci. 20, 303–330 65 Stein, J.F. (1986) Role of the cerebellum in the visual guidance of movement. Nature 323, 217–221 66 Holmes, G. (1917) The symptoms of acute cerebellar injuries due to gunshot injuries. Brain 40, 461–535 67 Kawato, M. et al. (1987) A hierarchical neural-network model for control and learning of voluntary movement. Biol. Cybern. 57, 169–185 68 Schweighofer, N. et al. (1998) Role of the cerebellum in reaching movements in humans. I. Distributed inverse dynamics control. Eur. J. Neurosci. 10, 86–94 69 Bastian, A.J. et al. (1996) Cerebellar ataxia: abnormal control of interaction torques across multiple joints. J. Neurophysiol. 76, 492–509 70 Miall, R.C. and Wolpert, D.M. (1996) Forwards models for physiological motor control. Neural Netw. 9, 1265–1279 71 Day, B.L. et al. (1998) Influence of vision on upper limb reaching movements in patients with cerebellar ataxia. Brain 121, 357–372 72 Johansson, R.S. and Westling, G. (1984) Roles of glabrous skin receptors and sensorimotor memory in automatic control of precision grip when rougher or more slippery. Exp. Brain Res. 56, 550–564 73 Muller, F. and Dichgans, J. (1994) Dyscoordination of pinch and lift forces during grasp in patients with cerebellar lesions. Exp. Brain Res. 101, 485–492 74 Martin, O. and Prablanc, C. (1991) Two-dimensional control of trajectories towards unconsciously detected double step targets. In Tutorials in Motor Neuroscience (Requin, J. and Stelmach, G.E., eds), pp. 581–598, Kluwer Desmurget and Grafton – Forward modeling 431 Trends in Cognitive Sciences – Vol. 4, No. 11, November 2000 Review