Reviews

Neuron, Vol. 46, 153–160, April 7, 2005, Copyright ©2005 by Elsevier Inc. DOI 10.1016/j.neuron.2005.02.009

Motor Learning by Observing

Andrew A.G. Mattar1,2,4 and Paul L. Gribble1,2,3,* 1Department of Psychology 2 Department of Physiology and Pharmacology 3 Graduate Program in Neuroscience University of Western Ontario London Ontario Canada N6A 5C2

Summary

Learning complex motor behaviors like riding a bicy- cle or swinging a golf club is based on acquiring neu- ral representations of the mechanical requirements of movement (e.g., coordinating muscle forces to con- trol the club). Here we provide evidence that mecha- nisms matching observation and action facilitate mo- tor learning. Subjects who observed a video depicting another person learning to reach in a novel mechani- cal environment (imposed by a robot arm) performed better when later tested in the same environment than subjects who observed similar movements but no learning; moreover, subjects who observed learning of a different environment performed worse. We show that this effect is not based on conscious strategies but instead depends on the implicit engagement of neural systems for movement planning and control.

Introduction

The human motor system can generate accurate move- ments under widely varying mechanical conditions. For example, a skilled athlete can accurately throw a light baseball or a heavy football, even though the underly- ing muscle forces are very different. This important fea- ture of the motor system is based on the acquisition of neural representations of the ways in which the envi- ronment’s mechanical properties affect the motor sys- tem (Brashers-Krug et al., 1996; Conditt et al., 1997; Flanagan and Wing, 1997; Gandolfo et al., 2000; Gribble and Scott, 2002; Shadmehr and Mussa-Ivaldi, 1994). Each mechanical context may be associated with a neural representation of its properties, which is used to specify the patterns of control signals to muscles that are required to generate an accurate movement in that context (Haruno et al., 2001; Wolpert and Kawato, 1998).

Recent advances in the understanding of motor learning have been based on experiments using robotic devices to create novel mechanical environments, which typically involve the application of forces that perturb the limb during movement. After an initial phase in which movements are perturbed from their intended trajectory, they eventually return to normal despite the ongoing application of forces to the arm. This adapta- tion (known as motor learning) is thought to reflect the

*Correspondence: pgribble@uwo.ca

4 Present address: Department of Psychology, McGill University, Montréal Québec Canada H3A 1B1.

acquisition of a neural representation of the novel me- chanical environment and its subsequent use by neural systems involved in limb control.

A powerful new idea in neuroscience links motor con- trol with action observation. When we observe the ac- tions of others, we activate the same neural circuitry responsible for planning and executing our own ac- tions. For example, so-called “mirror neurons” in the premotor cortex are activated both when observing an action and when performing the same action (Gallese et al., 1996; Rizzolatti et al., 1996). Evidence for a mech- anism linking observation and action has been demon- strated both in humans and nonhuman primates, in neurophysiological (Strafella and Paus, 2000; Watkins et al., 2003), brain-imaging (Buccino et al., 2001; Grafton et al., 1997; Iacoboni et al., 1999), and eye-tracking studies (Flanagan and Johansson, 2003). It has been proposed that this mechanism forms the basis by which we understand the actions of others (Carey, 1996; Rizzolatti et al., 2001; Wilson et al., 2004): by mapping a representation of observed actions onto motor systems, observers gain knowledge of those ac- tions by “internally” executing them. In a series of ex- periments, we test the intriguing possibility that such a system linking observation and action could facilitate motor learning.

It has been demonstrated that high-level information about the form of movements can be acquired by ob- serving the actions of others. For example, rats can learn the spatial relationships in a Morris water maze by observing other rats engaged in the same task (e.g., see Petrosini et al., 2003, for review). Studies examining reaction times indicate that human observers can learn finger-tapping sequences by watching others (Kelly et al., 2003). These experiments and others like them (Heyes and Foster, 2002; Vinter and Perruchet, 2002) show that information about “what” movements to make (details used at the planning stage, e.g., move- ment direction) can be acquired visually based on ob- servation. Here however, we address a new and funda- mentally different question: can information specifying “how” to make movements at the level of motor execu- tion (e.g., novel patterns of muscle forces) be conveyed through observation?

We used an experimental paradigm in which a ro- botic device generated novel force environments that perturbed the trajectory of the limb during reaching movements (Shadmehr and Mussa-Ivaldi, 1994). Our goal was to determine whether observing another indi- vidual undergoing the process of motor learning could affect the subsequent performance of naive observers. We show that neural representations of novel environ- ments can be acquired visually on the basis of observa- tion, and further experiments indicate that this process is not dependent on the use of conscious strategies but instead is based on the implicit engagement of motor systems. These findings broaden the scope of theories linking observation and action by demonstrating that by watching another individual learning to move, observers

Neuron 154

can learn not only what movements to make, but how s tto make them as well. b T

Results w p

Effects of Observing Motor Learning Subjects (n = 12) who observed a video depicting an- m other person learning a clockwise (CW) force field (FF) e (see Experimental Procedures and Figure 1) performed o significantly better when later tested in the same CWFF e than control subjects (n = 12) who did not observe but s rested for an equivalent amount of time (12 min). More- c over, another group of subjects (n = 12) who observed s learning of a counterclockwise (CCW) FF performed C worse than subjects who did not observe learning. Fig- C ure 2B shows examples of typical movement trajecto- t ries of subjects in each group as they first encountered the CWFF. Movement trajectories of subjects who ob- w served CWFF learning before encountering the CWFF t themselves were less curved than those of control sub- w jects and subjects who first observed CCWFF learning. b

Figure 3 shows mean learning curves for control sub- t jects and subjects who observed CW or CCWFF learn- t ing. Performance on each movement trial was esti- w mated by computing a measure of movement curvature m known as perpendicular distance (see Experimental t Procedures). All three groups of subjects reduced tra- t jectory curvature over time. When subjects who ob- C served CWFF learning first encountered the CWFF themselves, their movements were characterized by an a average of 23% less curvature than control subjects c

vwho observed nothing (Figure 3B). In contrast, when

Figure 1. Experimental Setup and Design

(A) Subjects sat grasping the end of the ro- botic device, which they used to guide an on-screen cursor to targets (see inset). (B) All subjects first performed 96 move- ments in a null field (no force field). In a first experiment, subjects were then randomly assigned to one of three groups who ob- served CWFF learning, CCWFF learning, or who observed nothing. All subjects were then tested in a CWFF. Subsequent control studies are described in Results.

ubjects who observed CCWFF learning first encoun- ered the CWFF, their movements were characterized y an average of 18% more curvature than controls. hus, while all subjects learned the CWFF, performance as significantly affected by having observed another erson learning CW or CCW force fields. Observation had immediate effects on subsequent otor performance (see Figure 4A). Significant differ-

nces were observed among mean curvature averaged ver the first eight movements (one to each target) in ach experimental condition (p < 0.001). Curvature for ubjects who observed CWFF learning was signifi- antly less than for control subjects (p < 0.05) and for ubjects who observed CCWFF learning (p < 0.01). onversely, curvature for subjects who observed CWFF learning was significantly greater than for con-

rols who observed nothing (p < 0.05). It should be noted that the force fields used here ere velocity dependent; thus, the magnitude of per-

urbing forces generated by the robot varied directly ith the speed of arm movement. To rule out the possi- ility that the observed differences in movement curva- ure were due to differences in the magnitude of per- urbing forces (due to differences in movement speed), e examined hand tangential velocity in each experi- ental group. No significant differences in peak hand

angential velocity (and hence the magnitude of per- urbing forces) were observed between the control, WFF, and CCWFF observation groups (p > 0.05). To assess potential differences in the temporal char-

cteristics of movement as a result of observation, we omputed four additional measures: time to peak cur- ature, time to peak difference in curvature (relative to

Motor Learning by Observing 155

Figure 2. Hand Trajectories

(A) Typical hand trajectories for movements to the eight targets in a NF (purple dashes) and initial exposure in the CWFF (solid blue lines). (B) Representative hand trajectories for movements to one target for subjects when first tested in a CWFF, after first observing noth- ing (blue), CWFF learning (green), or CCWFF learning (red). (C) Lateral displacement (X) plotted against time for the example trajectories shown in (B). (D) Lateral displacements (X) plotted relative to the trajectory for the control group (blue) who observed nothing prior to being tested in the CWFF. (E) Tangential velocity of the hand plotted against time for the ex- ample trajectories shown in (B).

Figure 3. Learning Curves

Mean learning curves for subjects when tested in a CWFF after first observing CWFF learning (green), CCWFF learning (red), or no observation (blue). (A) The curvature of hand trajectories is plotted against movement number. Each data point represents the average of eight movements. Vertical bars indicate 1 SEM. (B) Learning curves for the CWFF and CCWFF observation groups are plotted relative to the control group (horizontal dashed line at zero) and are expressed as a proportion of curvature for the first mean in the no observation group. Vertical bars indicate one SEM.

the no observation group), time to peak hand tangential velocity, and total movement time. MANOVA was used to test for effects of viewing condition. No significant effects of viewing condition were present for any of the four measures (p > 0.05 in all cases).

To rule out the possibility that subjects covertly moved their arm or activated arm muscles during ob- servation, we conducted a control experiment in which we recorded muscle activation patterns first during 96 movements in a null field (NF) and then during passive observation of the video depicting 96 trials of CWFF learning. Surface electrodes were used to record mus- cle activation patterns from four shoulder and elbow muscles (see Experimental Procedures).

For recordings made while subjects performed movements in a NF, typical biphasic and triphasic pat- terns of agonist and antagonist muscle activity were seen in all four muscles and for all eight movement di- rections. Two-factor repeated measures MANOVA and Tukey post hoc tests were used to test for differences between mean EMG as a function of movement direc- tion, across three time windows: a baseline, agonist, and antagonist window (see Experimental Procedures). Significant differences between mean EMG in baseline

versus agonist and antagonist windows were seen in all eight movement directions for pectoralis and deltoid (p < 0.01) and in six out of eight movement directions for biceps and triceps (p < 0.01). The particular pattern of differences depended on movement direction (Hasan and Karst, 1989; Karst and Hasan, 1991). For record- ings made during passive observation, no visibly de- tectable muscle activation patterns were seen for any subject during any point in the observation session. Nevertheless, to quantitatively test for the possibility of muscle activations, we again used MANOVA to test differences between baseline and agonist and baseline and antagonist EMG as a function of target direction. No significant differences were detected between baseline EMG and agonist or antagonist EMG for any muscle or any target direction (p > 0.05 in all cases).

It is possible that the effects of observing CW and CCWFF learning on subsequent performance may have been due in part to some nonspecific effect of observ- ing curved hand motions rather than the observation of motor learning itself. To rule this out, we tested another

Neuron 156

R W w p t t n t l s o f s m t v w a t i t t i a a

Figure 4. Trajectory Curvature as a Function of Observation m (A) Mean trajectory curvature averaged over the first eight move- ments in the CWFF for subjects who first observed CWFF learning p (CW), CCWFF learning (CCW), or no observation (no OBS). Vertical

tbars indicate one SEM. *p < 0.05; **p < 0.01. t(B) Effect of observation expressed as the decrease in curvature srelative to the control group who observed nothing for subjects

who observed movements in a random FF (Random), for subjects d who observed CWFF learning (CW), for subjects who observed g CWFF learning while performing a distraction task (Distraction

n Task), and while performing rhythmic arm movements (Motor En-

fgagement). Vertical bars indicate one SEM. *p < 0.05. f t

a

group of subjects (n = 12) who observed an individual dattempting to learn a randomly varying FF. The FF pre- osented by the robot was randomly varied from trial to ttrial between a CW, CCW, or null field. Thus, subjects iwho observed an individual attempting to learn the ran- Cdom FF were exposed to the same kinds of hand mo-

tions as in the original CW or CCW FF conditions, but odid not observe the progressive and systematic de- ccrease in movement curvature over time typically asso- vciated with motor learning. Figure 4B shows the perfor- Wmance of subjects who observed a random FF, vaveraged over the first eight trials when first exposed yto the CWFF. Performance is plotted as the decrease tin curvature relative to controls who observed noth- oing—thus, values near zero indicate little benefit of ob- wservation, while large values indicate a large benefit of pobservation (a large decrease in curvature). The de- pcrease in curvature for subjects who observed the ran- tdom FF was not significantly different than zero (p > p0.05). This indicates that performance in a CWFF was mnot influenced by the observation of a random FF. For ccomparison, the significant decrease in curvature as a lresult of observing CWFF learning is plotted on the

same scale (p < 0.01). g

ole of Conscious Strategies e used a distractor paradigm to assess the extent to hich the effect of observation on subsequent motor erformance depends on conscious strategy forma- ion. A separate group of subjects (n = 12) was asked o perform an arithmetic addition task while simulta- eously observing learning of a CWFF. Beside each of he movement targets in the video depicting CWFF earning, a number between 1 and 8 appeared as the ubject in the video began to move. The task for the bserver was to add the current number to the number rom the previous movement and verbally indicate the um to the experimenter. Subjects were not required to onitor the speed of observed movements (unlike in

he original experiment). Thus, the distractor task in- olved both an arithmetic operation and a load on orking memory. The task was designed in such a way s to be easy enough to perform while still observing he motor learning depicted in the videos, but challeng- ng enough that subjects required attention and cogni- ive effort to complete the task. This task is similar to hose that frequently appear in the cognitive literature n divided-attention paradigms (Baddeley, 2003; Tulving nd Craik, 2000). The proportion of errors during the rithmetic/memory task ranged across subjects to a aximum of 8.9% (mean = 2.1%, SD = 1.5%). After observing CWFF learning and simultaneously

erforming the distractor task, subjects were exposed o the CWFF. Subjects in the distractor group showed he same benefit from observing CWFF learning as ubjects in the original experiment. Figure 4B plots the ecrease in curvature for the subjects in the distractor roup relative to the control subjects who observed othing, averaged over the first eight movements when irst exposed to the CWFF. The decrease in curvature or the distractor group (expressed as a proportion of he curvature in the control group who did not observe nything, mean = 21.4% ± 4.9%) was not significantly ifferent from that of the original group of subjects who bserved CWFF learning (p > 0.05). Thus, the distrac- ion task did not reduce the beneficial effect of observ- ng CWFF learning on subsequent performance in a WFF. To further assess the extent to which the effect of

bserving motor learning may have been due to the onscious formation of movement strategies, we inter- iewed subjects after the end of exposure to the CWFF. e asked subjects in the CW, CCW, and random obser-

ation groups the following question: “were the forces ou felt when you were moving the robot the same as he forces that were shown in the video?” The number f correct responses (14 correct responses out of 36) as not significantly different than what would be ex- ected if subjects were randomly guessing (χ2 analysis, > 0.05). Subjects were not aware of how the forces

hey experienced in the CWFF related to the forces de- icted in the CW or CCWFF recordings. Thus, although otor performance of subjects in the CWFF was signifi-

antly affected by the observation of CW and CCWFF earning, this effect was not based on conscious strate- ies (e.g., “I should try to push to the left”).

Motor Learning by Observing 157

Role of Motor Systems We used a paradigm involving the performance of unre- lated arm movements to assess the extent to which the effect of observation is based on the activation of systems for motor control. A group of subjects (n = 12) was instructed to slowly move their arm in a circular motion while observing another person learning a CWFF. To eliminate any systematic bias due to the di- rection of circular motions, subjects were instructed by the experimenter to alternate movements between CW and CCW directions once, halfway through the obser- vation session. After observation, subjects were tested in the CWFF. The beneficial effect of observing CWFF learning was significantly reduced for subjects in this “motor engagement” group. Figure 4B shows the mean reduction in curvature for the motor engagement group plotted beside the mean reduction in curvature for the original CWFF observation group and the attentional distraction group. For subjects who moved their arm while observing CWFF learning, the reduction in move- ment curvature when first exposed to the CWFF (mean reduction = 10.4% ± 4.6%) was significantly less than for subjects who did not move their arm (p < 0.05). The reduction in curvature was still significantly greater than zero (p < 0.05), indicating that although the magni- tude of the beneficial effect was reduced, subjects still received some benefit from observing CWFF learning.

To control for the possibility that the observed de- crease in performance in the CWFF may have resulted from factors related to moving one’s arm in a circular pattern (e.g., fatigue or unintended motor learning), we tested an additional group (n = 12) who were asked to perform the same motions for 12 min prior to being tested in a CWFF. This group did not observe anything during these 12 min but only performed the circular arm movements. These subjects performed no differently when tested in the CWFF than controls who did not perform circular arm movements. Mean curvature dur- ing the first eight movements was not significantly dif- ferent than for control subjects (p > 0.05). Thus, the performance of circular arm movements on its own had no effect on subsequent performance in a CWFF.

Discussion

Here we have shown that by observing another indivi- dual learning to move accurately in a novel mechanical environment, observers move more accurately them- selves. Subjects can acquire neural representations of novel force environments on the basis of visual infor- mation. Further, while motor learning by observing does not depend on conscious awareness of the observer, the tendency for an unrelated movement task to signifi- cantly reduce the ability of subjects to learn by observ- ing indicates that the implicit engagement of motor systems is required.

Other work has shown that information used in plan- ning movement can be acquired via observation. These studies have demonstrated that kinematic (spatio-tem- poral) information specifying figural aspects of move- ment (e.g., “what” movements to make) can be con- veyed visually (Heyes and Foster, 2002; Kelly et al.,

2003; Vinter and Perruchet, 2002). The present results are quite different and represent experimental evidence that observers can extract information used at the level of motor execution (e.g., “how” to make movements) on the basis of observation. By observing another indi- vidual learning to move accurately in a novel force envi- ronment, the observer was able to form a neural repre- sentation of the environment’s mechanical properties, which was subsequently put to use in controlling move- ments in the CWFF.

Performance in the CWFF varied depending on what subjects had previously observed. Observation of a CWFF facilitated later performance in the same CWFF, while observation of a CCWFF disrupted performance in the CWFF. These findings are consistent with the idea that, as a result of observation, subjects were able to predict the influence of the observed FF on the arm. Acquired representations of the FFs, rather than non- specific strategies (e.g., muscle co-contraction) gov- erned movement.

The finding that subjects can learn something useful about novel force environments on the basis of obser- vation is remarkable, given the complex relationship between movement kinematics and associated time- varying neural control signals to muscles. As a subject observes another person moving in a novel force envi- ronment (e.g., a CWFF), the only information directly available to the observer is visual in nature and speci- fies kinematic aspects of movement. In order for an ob- server to learn something about a novel mechanical en- vironment, the nervous system must first assume that any deviations from a typical straight-line hand trajec- tory (Morasso, 1981) represent movement errors. On the basis of these errors, the motor system must then construct a representation of the perturbing forces that resulted in the observed hand trajectory. This would re- quire an implicit model of the mechanical characteris- tics (e.g., stiffness) of the limb and its predicted re- sponse to external forces. Finally, in order to benefit from this learning, the motor system must determine the changes in neural control signals to muscles that would be required in order to oppose the predicted per- turbing forces. Information about movement kinemat- ics, acquired from visual information alone, must be transformed into a representation of forces and subse- quently the required changes to neural control signals for movement. It is likely that the neural bases of motor learning by observing share the same substrates that have been described for sensorimotor transformations in overt voluntary movement (Cohen and Andersen, 2002; Kakei et al., 2003; Kalaska et al., 1997; Snyder, 2000).

We used a distraction task to determine the role of explicit, conscious strategies in motor learning by ob- serving. Our findings indicate that observers can bene- fit from observation even when attentional systems are engaged by a distractor task, suggesting that these systems are not critical for motor learning by observing. While it could be argued that attentional or cognitive systems are indeed involved and that our distraction task simply failed to engage these systems to an ade- quate extent, this seems unlikely. Our subjects commit- ted errors on the distraction task (see Results), indicat-

Neuron 158

ing that it was challenging and required effort to C complete. Error rates indicated that subjects were not t ignoring the distraction task and simply attending to m the motor aspects of the video recording. Another pos- w sibility is that, as an arithmetic task, the distractor only v engaged mathematical and working memory systems, s leaving other attentional mechanisms free to form con- s scious, explicit strategies. We interviewed subjects fol- t lowing testing in the CWFF to determine whether they used strategies during observation. Subjects could not c correctly identify whether the FF observed was the t same or different than the FF experienced (see Re- p sults), suggesting that subjects did not use explicit r strategies to guide their performance. This lack of a dependence on explicit strategies has an intriguing im-

E plication, namely that motor learning by observing may occur unbeknownst to the subject. S

In contrast, motor learning by observing was com- 8 promised when the motor system was engaged with t

can unrelated movement task. The beneficial effect of Uobserving motor learning was reduced in subjects who wperformed rhythmic arm movements during observa- s

tion (see Results). This suggests that motor systems m

are involved in acquiring neural representations of novel environments during observation. Recent find-

Rings indicate that observation of movement activates Smotor areas involved in producing the same movement T

(see Rizzolatti et al., 2001, for review). Our results sug- t

gest that the ability of visual information to drive motor ( learning through systems linking observation and ac- s

wtion is significantly diminished when motor systems are toccupied by the generation of unrelated movements. tIt should be noted that not all observed motor beha- tvior leads to motor learning in observers. We have

shown here that observing an individual experiencing v a randomly varying mechanical environment does not ( affect the subsequent performance of observers. For a

amotor learning by observing to occur, the observer dmust presumably be exposed to systematic movement (errors so that a representation of perturbing forces may r

be developed. Indeed, recent work has shown that mo- w

tor cortical areas are activated when subjects observe t movement error (van Schie et al., 2004). Presumably, g observing the actions of skilled individuals (after learn- m ing has already occurred) would not lead to motor

blearning in the observer. rIt is important to consider what specific information a

may be required for motor learning by observing. Pre- d

sumably, consistent information about the nature of perturbing forces is required. In the present study, sub- jects observing CWFF and CCWFF learning were ex- posed to consistent CW or CCW movement errors that gradually decreased over time. Whether motor learning by observing depends on exposure to a gradual de-

w crease in movement error over time cannot be ad- f dressed in the present study. However, it seems likely t that observers would still benefit from exposure to sys- a

etematic movement errors that do not decrease over Atime. (While performance in the CWFF clearly benefited

from the prior observation of CWFF learning, subjects w

in the present study were not able to fully learn how to R move accurately in the CWFF solely from observation. v

oAdditional experience performing movements in the

WFF was required to further reduce movement curva- ure. Nevertheless, considerable changes in perfor- ance were seen as a result of observation. Subjects ho observed CWFF learning gained a significant ad- antage over control subjects who did not. Similarly, ubjects who observed CCWFF learning experienced a ignificant and longer-lasting disadvantage compared o control subjects.

In summary, we have shown that motor learning oc- urs in the absence of overt movement by observing he actions of others. The human motor system incor- orates the experiences of others in building the motor

epertoire of the individual.

xperimental Procedures

ubjects 4 subjects (mean age 21.02 ± 0.39 SE, 40 males) participated in he experiments described here. All subjects provided informed onsent to procedures that complied with guidelines set out by the niversity of Western Ontario’s Research Ethics Board. All subjects ere right-hand dominant for writing, had normal or corrected vi- ion and reported no neurological or musculo-skeletal impair- ents.

obotic Device ubjects used the InMotion2 robotic device (Interactive Motion echnologies) to guide an on-screen cursor to a series of visual argets presented using a system of mirrors and an LCD projector see Figure 1A). Subjects sat in front of a custom-designed tabletop urface with their right arm supported by a padded air-sled, which as connected to a compressed air source to provide virtually fric-

ionless motion and supported the arm against gravity. The level of he chair was adjusted so that the shoulder was abducted 90° from he sagittal plane.

Visual targets were presented to the subject using a semi-sil- ered mirror placed between the arm and a back-projection screen see Figure 1A). Targets thus appeared to “float” in the same plane s the hand. A total of 8 movement targets were used, placed equ- lly around the circumference of a circle. Targets were 24 mm in iameter and were located 10 cm away from a central start location

see Figure 1). Subjects were instructed to move quickly and accu- ately to targets in a single continuous motion. Movement speed as controlled by providing subjects with feedback on a trial-to-

rial basis. The color of the target changed to blue (correct speed), reen (too slow) or red (too fast) according to the measured move- ent speed on each trial. Desired movement duration was 375 ms. The robot was programmed to alter the dynamics of limb motion

y applying forces (“force fields”, FF) to a subject’s arm during eaching movements to targets. Forces were velocity dependent nd were applied in a clockwise (CW) or counterclockwise (CCW) irection according to the following equation:

[FxFy] = [ 0 dk −dk 0 ][

ẋ ẏ] (1)

here Fx and Fy are robot-generated forces in the left/right and orward/backward direction, respectively, ẋ and ẏ are hand veloci- ies, k = 20 Ns/m, and d = +1.0 (CW) or −1.0 (CCW). Thus, forces pplied by the robot were zero at movement start and movement nd and reached a maximum at peak hand tangential velocity. cross subjects, the mean peak force applied to the arm was 4.9 N

SD = 0.9 N). Robot forces were controlled using custom software routines ritten in C and Tcl programming languages and run within the T Linux operating system on a Pentium 4 CPU. Robot positions, elocities, and applied forces were sampled at 200 Hz and stored n a digital computer for offline analysis.

Motor Learning by Observing 159

Video Recordings Video recordings provided subjects with a top-down view of an- other individual’s right arm and the workspace within which move- ments to targets were made. Superimposed on the image of the arm were the visual targets and a cursor representing the position of the hand (see Figure 1B). Recordings were made using a digital video camera and were edited using Final Cut Pro 4 software (Ap- ple Computer). Each recording was approximately 6 min in duration and demonstrated a series of 96 movements. Subjects were shown the appropriate video twice.

The recordings depicted an individual moving to targets as the robot applied perturbing forces to the arm. In the CWFF recording, forces were the same as those later experienced by the observer; in the CCWFF recording, the forces were applied in the opposite direction. These recordings showed the progression from highly perturbed to straight movements typically associated with motor learning (e.g., Figure 3A).

The random FF recording showed an individual interacting with the robotic device as it generated randomly varying perturbing forces. Subjects of course were not able to learn such an environ- ment (Takahashi et al., 2001). Thus, the video demonstrated mo- tions that were similar to those in the CW and CCWFF recordings but which lacked the progression from perturbed to straight move- ments associated with motor learning.

In the video used in the distraction condition, a digit from 1 through 8 was superimposed onto the CWFF recording at each target location. Subjects were asked to sum the digits indicated by successive movements (current + previous) and to indicate the re- sult verbally to the experimenter.

Instructions to Subjects Subjects were asked to use the robotic device to guide a cursor to targets. Following their initial familiarization with the task (96 movements with no forces applied, see Figure 1B), subjects were asked to observe a video recording of another individual perform- ing a similar task. No mention was made of the CW, CCW, or ran- dom FFs depicted in the recordings. To ensure that subjects paid attention to the video recordings, we asked them to monitor the depicted movements and report to the experimenter when move- ments made by the subject in the video were too fast or slow (this was indicated by the targets changing color). Subjects were highly accurate in this regard (mean score > 98% correct). During obser- vation, subjects were instructed to let go of the robot handle and to rest their arm on the tabletop surface. Following observation, subjects were again asked to guide the cursor to targets. Subjects were not warned that the robot would apply a CWFF. At the com- pletion of the experiment, subjects were questioned with respect to their awareness of the FFs observed and experienced.

EMG Recordings Electromyographic signals (EMG) were recorded from biceps long head, triceps lateral head, pectoralis clavicular head, and posterior deltoid using surface electrodes (Delsys). Signals were sampled at 1000 Hz, band-pass filtered between 30–300 Hz, and rectified prior to analysis. Mean EMG was computed during three windows time- aligned to movement onset in the NF movements and time-aligned to the onset of each movement in the video depicting CWFF learn- ing. An initial 200 ms baseline window beginning 300 ms prior to movement onset was used to characterize baseline levels of EMG. An agonist window beginning 100 ms prior to movement onset and ending 100 ms after movement onset was used to characterize phasic agonist muscle activation associated with movement accel- eration. An antagonist window beginning 150 ms after movement onset and lasting 200 ms was used to characterize antagonist mus- cle activity associated with movement deceleration. Five subjects were tested in the control study.

Measures and Statistics Performance on each movement trial when subjects were tested in the CWFF was quantified using a measure of movement curvature defined as the maximum perpendicular deviation from a line seg- ment linking movement start position and the target’s location (Malfait et al., 2002; Shadmehr and Brashers-Krug, 1997; Thor-

oughman and Shadmehr, 1999). Other similar measures such as angular error and path length yielded qualitatively similar results. Individual scores were collapsed across bins of eight movements, and differences between group means were tested using multivari- ate analyses of variance (MANOVAs) and Tukey post hoc tests. Data analyses were carried out using custom software routines written using Matlab (The Mathworks).

Acknowledgments

The authors thank D. Debicki, N. Cothros, L. Brown, D. Shiller, and D. Ostry for helpful comments on early versions of the manuscript and S. Köhler for fruitful discussions about experimental design. Research was supported by CIHR (Canada). The authors declare that they have no competing financial interests.

Received: October 8, 2004 Revised: January 4, 2005 Accepted: February 7, 2005 Published: April 6, 2005

References

Baddeley, A. (2003). Working memory: looking back and looking forward. Nat. Rev. Neurosci. 4, 829–839.

Brashers-Krug, T., Shadmehr, R., and Bizzi, E. (1996). Consolidation in human motor memory. Nature 382, 252–255.

Buccino, G., Binkofski, F., Fink, G.R., Fadiga, L., Fogassi, L., Gal- lese, V., Seitz, R.J., Zilles, K., Rizzolatti, G., and Freund, H.J. (2001). Action observation activates premotor and parietal areas in a so- matotopic manner: an fMRI study. Eur. J. Neurosci. 13, 400–404.

Carey, D.P. (1996). ‘Monkey see, monkey do’ cells. Neurophysiol- ogy. Curr. Biol. 6, 1087–1088.

Cohen, Y.E., and Andersen, R.A. (2002). A common reference frame for movement plans in the posterior parietal cortex. Nat. Rev. Neu- rosci. 3, 553–562.

Conditt, M.A., Gandolfo, F., and Mussa-Ivaldi, F.A. (1997). The mo- tor system does not learn the dynamics of the arm by rote memori- zation of past experience. J. Neurophysiol. 78, 554–560.

Flanagan, J.R., and Johansson, R.S. (2003). Action plans used in action observation. Nature 424, 769–771.

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.

Gallese, V., Fadiga, L., Fogassi, L., and Rizzolatti, G. (1996). Action recognition in the premotor cortex. Brain 119, 593–609.

Gandolfo, F., Li, C., Benda, B.J., Schioppa, C.P., and Bizzi, E. (2000). Cortical correlates of learning in monkeys adapting to a new dy- namical environment. Proc. Natl. Acad. Sci. USA 97, 2259–2263.

Grafton, S.T., Fadiga, L., Arbib, M.A., and Rizzolatti, G. (1997). Pre- motor cortex activation during observation and naming of familiar tools. Neuroimage 6, 231–236.

Gribble, P.L., and Scott, S.H. (2002). Overlap of internal models in motor cortex for mechanical loads during reaching. Nature 417, 938–941.

Haruno, M., Wolpert, D.M., and Kawato, M. (2001). Mosaic model for sensorimotor learning and control. Neural Comput. 13, 2201– 2220.

Hasan, Z., and Karst, G.M. (1989). Muscle activity for initiation of planar, two-joint arm movements in different directions. Exp. Brain Res. 76, 651–655.

Heyes, C.M., and Foster, C.L. (2002). Motor learning by observa- tion: evidence from a serial reaction time task. Q. J. Exp. Psychol. A 55, 593–607.

Iacoboni, M., Woods, R.P., Brass, M., Bekkering, H., Mazziotta, J.C., and Rizzolatti, G. (1999). Cortical mechanisms of human imita- tion. Science 286, 2526–2528.

Neuron 160

Kakei, S., Hoffman, D.S., and Strick, P.L. (2003). Sensorimotor transformations in cortical motor areas. Neurosci. Res. 46, 1–10.

Kalaska, J.F., Scott, S.H., Cisek, P., and Sergio, L.E. (1997). Cortical control of reaching movements. Curr. Opin. Neurobiol. 7, 849–859.

Karst, G.M., and Hasan, Z. (1991). Timing and magnitude of electro- myographic activity for two-joint arm movements in different direc- tions. J. Neurophysiol. 66, 1594–1604.

Kelly, S.W., Burton, A.M., Riedel, B., and Lynch, E. (2003). Sequence learning by action and observation: evidence for separate mecha- nisms. Br. J. Psychol. 94, 355–372.

Malfait, N., Shiller, D.M., and Ostry, D.J. (2002). Transfer of motor learning across arm configurations. J. Neurosci. 22, 9656–9660.

Morasso, P. (1981). Spatial control of arm movements. Exp. Brain Res. 42, 223–227.

Petrosini, L., Graziano, A., Mandolesi, L., Neri, P., Molinari, M., and Leggio, M.G. (2003). Watch how to do it! New advances in learning by observation. Brain Res. Brain Res. Rev. 42, 252–264.

Rizzolatti, G., Fadiga, L., Gallese, V., and Fogassi, L. (1996). Premo- tor cortex and the recognition of motor actions. Brain Res. Cogn. Brain Res. 3, 131–141.

Rizzolatti, G., Fogassi, L., and Gallese, V. (2001). Neurophysiologi- cal mechanisms underlying the understanding and imitation of ac- tion. Nat. Rev. Neurosci. 2, 661–670.

Shadmehr, R., and Brashers-Krug, T. (1997). Functional stages in the formation of human long-term motor memory. J. Neurosci. 17, 409–419.

Shadmehr, R., and Mussa-Ivaldi, F.A. (1994). Adaptive representa- tion of dynamics during learning of a motor task. J. Neurosci. 14, 3208–3224.

Snyder, L.H. (2000). Coordinate transformations for eye and arm movements in the brain. Curr. Opin. Neurobiol. 10, 747–754.

Strafella, A.P., and Paus, T. (2000). Modulation of cortical excitabil- ity during action observation: a transcranial magnetic stimulation study. Neuroreport 11, 2289–2292.

Takahashi, C.D., Scheidt, R.A., and Reinkensmeyer, D.J. (2001). Im- pedance control and internal model formation when reaching in a randomly varying dynamical environment. J. Neurophysiol. 86, 1047–1051.

Thoroughman, K.A., and Shadmehr, R. (1999). Electromyographic correlates of learning an internal model of reaching movements. J. Neurosci. 19, 8573–8588.

Tulving, E., and Craik, F.I.M. (2000). The Oxford Handbook of Mem- ory (London: Oxford University Press).

van Schie, H.T., Mars, R.B., Coles, M.G., and Bekkering, H. (2004). Modulation of activity in medial frontal and motor cortices during error observation. Nat. Neurosci. 7, 549–554.

Vinter, A., and Perruchet, P. (2002). Implicit motor learning through observational training in adults and children. Mem. Cognit. 30, 256–261.

Watkins, K.E., Strafella, A.P., and Paus, T. (2003). Seeing and hear- ing speech excites the motor system involved in speech prod- uction. Neuropsychologia 41, 989–994.

Wilson, S.M., Saygin, A.P., Sereno, M.I., and Iacoboni, M. (2004). Listening to speech activates motor areas involved in speech prod- uction. Nat. Neurosci. 7, 701–702.

Wolpert, D.M., and Kawato, M. (1998). Multiple paired forward and inverse models for motor control. Neural Netw. 11, 1317–1329.