Behavioural Brain Research 278 (2015) 569576

Contents lists available at ScienceDirect

Behavioural Brain Research

journal homepage: www.elsevier.com/locate/bbr

Research report

Sub-processes of motor learning revealed by a robotic manipulandum

for rodents
O. Lambercy a,1 , M. Schubring-Giese b,1 , B. Vigaru a , R. Gassert a , A.R. Luft b,c, , J.A. Hosp b
a
b
c

Rehabilitation Engineering Laboratory, Department of Health Sciences and Technology, ETH Zurich, LEO B9.1 Leonhardstrasse 27, 8092, Zurich, Switzerland
Vascular Neurology and Neurorehabilitation, Department of Neurology, University of Zurich, Frauenklinikstrasse 26, 8032, Zurich, Switzerland
Department of Neurology, Johns Hopkins University, 1550 Orleans Street, Baltimore MD 21231, USA

h i g h l i g h t s
We examine motor learning (easy vs. complex task) using a robotic device for rats.
Both tasks differ in performance level and the temporal evolution of kinematic parameters.
Different sub-processes of motor learning can be revealed in the complex task.

a r t i c l e

i n f o

Article history:
Received 12 June 2014
Received in revised form 22 October 2014
Accepted 30 October 2014
Available online 8 November 2014
Keywords:
Motor learning
Robot
Rodent
ETH Pattus

a b s t r a c t
Rodent models are widely used to investigate neural changes in response to motor learning. Usually,
the behavioral readout of motor learning tasks used for this purpose is restricted to a binary measure
of performance (i.e. successful movement vs. failure). Thus, the assignability of research in rodents
to concepts gained in human research implying diverse internal models that constitute motor learning is still limited. To solve this problem, we recently introduced a three-degree-of-freedom robotic
platform designed for rats (the ETH-Pattus) that combines an accurate behavioral readout (in the form
of kinematics) with the possibility to invasively assess learning related changes within the brain (e.g. by
performing immunohistochemistry or electrophysiology in acute slice preparations).
Here, we validate this platform as a tool to study motor learning by establishing two forelimb-reaching
paradigms that differ in degree of skill. Both conditions can be precisely differentiated in terms of their
temporal pattern and performance levels. Based on behavioral data, we hypothesize the presence of
several sub-processes contributing to motor learning. These share close similarities with concepts gained
in humans or primates.
2014 Elsevier B.V. All rights reserved.

1. Introduction
Motor learning as a form of procedural learning [1] is generally dened as the gradual improvement of motor performance
with practice [2]. Studies in primates and humans revealed several different sub-processes that contribute to this improvement
in a specic sequence: an initial phase of dening basic movement strategies [3] is followed by a period of rening movement
precision, whereas movement speed increases in a nal phase [4,5].

Corresponding author at: Division of Vascular Neurology and Neurorehabilitation, Department of Neurology, University Hospital of Zurich, Frauenklinikstrasse
26, 8091 Zurich, Switzerland. Tel.: +41 44 255 5400; fax: +41 44 255 4649.
E-mail addresses: andreas.luft@uzh.ch, andreas.luft@usz.ch (L. A.R.).
1
At equal contribution.
http://dx.doi.org/10.1016/j.bbr.2014.10.047
0166-4328/ 2014 Elsevier B.V. All rights reserved.

Animal models of motor skill learning involve tasks that are

not familiar to the animal, e.g. skilled forelimb reaching [6]. However, skilled reaching is restricted to a binary performance-measure
(i.e. did the movements meet the requirements dened by the
experimenter or not?) [7]. Only by observation or video tracking,
limited information about movement strategies and kinematics can
be obtained. It is therefore not possible to identify sub-processes of
motor learning as it was done in human experiments [8,9].
Recently we introduced the ETH Pattus [10,11], a robotic platform designed for interaction with forelimb movements in rats that
allows for automatized training and accurate behavioral readout in
the form of interaction kinematics. Here, we investigate two different reaching paradigms that differ in complexity, i.e. a 10 mm
free-pulling task (FP) and a 10 mm straight-pulling task (SP) where
an additional precision constraint is added for the movement to
be considered successful. Based on kinematic data acquired with

570

O. Lambercy et al. / Behavioural Brain Research 278 (2015) 569576

the robot, we report on the evolution of sub-processes constituting

motor learning under both training paradigms.
2. Materials and methods
2.1. Animals and experiments
Nave adult 1012 weeks old male Long-Evan rats (n = 19;
220270 g; Center dElevage R. Janvier, Le GenestSt. Isle, France)
were used for this study. Animals were housed in cages in groups
of three individuals in a 12/12-h light/dark cycle (light on: 8 pm,
off: 8 am). Animals were food-deprived for 24 h prior to the rst
training session. Daily food supplements (ca. 50 g/kg of standard
diet) were given after training to maintain constant body weight.
Access to water was ad libitum. All experiments were conducted
in accordance with Swiss regulations and were approved by the
Committee for Animal Experimentation of the Canton of Zrich.
2.2. Experimental setup
The ETH Pattus (Fig. 1A and B, for more detailed information
[10,11]) was used to provide kinematic analysis of movements
during acquisition of different motor tasks. The ETH Pattus is a
three-degree-of-freedom robotic manipulandum that allows planar movement in xy-direction and pro/supinationalthough the
tasks designed for this study required merely movements within
the xy-plane. It was designed to particularly meet the kinematic
requirements of rat forelimb movements (e.g. planar reaching,
grasping and pulling). Rats interact with the robot via a spherical handle (end-effector, 6 mm diameter) that can be manipulated
in order to perform a specic motor task, which is automatically
rewarded when accomplished correctly. During movements, the
handle position and velocity are continuously recorded with a
sampling frequency of 1 kHz. The ETH Pattus is placed in front of a
custom-made Plexiglas chamber (width: 400 mm, depth: 150 mm,
height: 450 mm) with a vertical opening (width: 10 mm, height:
50 mm). The handle is located 55 mm above the ground. In the back
of the chamber, a tray is mounted to hold pellets (45 mg, Bioserve
Inc., Frenchtown, NJ, USA) which are delivered by a pellet dispenser
(Model 80208, Lafayette Instrument Comp., IN, USA) in case of a
successful trial.
2.3. Behavioral conditions
The behavioral protocol consisted of three different phases: (1)
to familiarize the animals with the new environmental conditions
(without the robot), rats were placed in the Plexiglas chamber for
1 h/day whereas a quantity of 2030 pellets was freely administered. In the majority of cases animals consumed all pellets after
12 days and were subsequently assigned to the second phase of
training the next day. (2) To familiarize the animals to the robot
and especially to the end-effector, the second phase consisted of
positioning the robot handle in close distance (4 mm) to the chamber window. When animals touched the handle and displaced it
by a minimum of 0.2 mm in the x- or y-direction, an auditory cue
(beep sound, 1 s) was presented and rats were rewarded with a
food pellet delivered into the tray at the rear wall of the training chamber. When task performance reached 200 successful trials
in 60 min, animals started motor skill training. Repeated measurement ANOVA showed no statistically signicant difference in
performance level (i.e. number of trials completed per minute)
between groups (F = 1.29; p = 0.28) emphasizing a comparable baseline performance prior to the start of phase 3. (3) To allow 10 mm
pulling movements and to avoid the end-effector from entering
the chamber window at the end of the movement, training was
initiated by increasing the distance between the handle and the

chamber window from 4 mm to 18 mm. Animals were then split

into two groups and trained on either of two motor-learning tasks.
For both tasks, animals had to perform 200 trials a day (successful
or not). Thus, this number was constant over all animals, groups
and over time.
Free-pulling group (FP; n = 6; Fig. 1C): Animals had to reach out for
the handle and pull it over a distance of at least 10 mm in y-direction
within an area of 12 mm distance from the x-axis corresponding to the workspace limits of the robot. When animals fullled
this criterion, a trial was rated as success and an auditory cue
(beep sound, 1 s) was presented, while a pellet was automatically
released into the tray. At the end of each trial the handle was automatically retracted outside of the reachable workspace of the rat,
before moving back in front of the chamber window to initiate the
next trial.
Straight-pulling group (SP; n = 13; Fig. 1C): To increase the degree
of difculty, a precision constraint was added, requiring the rat to
pull the handle within a corridor around the straight line (i.e. y-axis)
of 2 mm in width to either side (x-direction).
Motor training sessions (phase 3) consisted of 200 trials. If the
animals did not perform 200 trials, a session was terminated after
60 min. Overall, training sessions were conducted over 25 consecutive days, with one session per day. Particular attention was paid
with respect to the handedness of a rat. Once an animal developed
a preference for a paw, it maintained this preference throughout
the experiment. We therefore refrained from reinforcing a certain
forelimb.
2.4. Data processing
During all trials of phase 3, the position (x, y) of the handle was
recorded by the ETH Pattus at a sampling frequency of 1 kHz and
stored on a desktop computer. Ofine data processing and analysis
was performed using Matlab (Matworks Inc., Natick MA, USA; Supplementary Fig. 1). Velocity signals were low-pass ltered using a
2nd order Butterworth lter with a cutoff frequency of 50 Hz. For
each trial, the initiation of the rats pulling movement was then
determined using an empirically chosen velocity threshold along
the pulling direction. This threshold (vy > 30 mm/s) was dened
rather high to differentiate the pulling movement from unspecic
displacement that occurs when rats initially grab the handle. A trial
was automatically rated as fail when the handle was released
before the pulling-distance of 10 mm was completed, when the
pulling movement was interrupted for more than 2 s or when the
handle was not touched for >180 s. For the analysis of kinematics
and parameter extraction, all valid (i.e. no fail) trials were taken
into account for both, SP and FP group regardless if they were rated
as successful or not. Position- and velocity- (norm of the velocity
vector) traces were further resampled using b-spline interpolation
to obtain a constant number of samples per trial and allow for calculation of mean trajectories and mean velocity proles over an
entire session.
For comparison with a well-established measure of skill learning [7,12] percentages of successful trials per session (success rate;
inter-session learning; SR), as well as per blocks of quintiles (=40
trials) within each session (intra-session learning) were calculated.
As a measure of motivation and handling of the operant conditioning paradigm incorporated into the task, the grasping latencies
(L) between the automatic positioning of the robot in front of the
chamber window and the grasp of the end-effector by the animal
were calculated. Four additional parameters were extracted from
the kinematic data collected by the ETH Pattus: (1) mean variability
in trajectories (VAR) within a session was evaluated by calculating
the area of the 95% condence interval of the mean of all successful pulling movements. This parameter is thought to display
the formation of a strategy for successful movements [13]. (2) As

O. Lambercy et al. / Behavioural Brain Research 278 (2015) 569576

571

an approximation of submovements that compose a movement

sequence, the number of zero crossings of the acceleration prole
(ZCA) was computed to gain insight into the development of movement smoothness (Supplementary Fig. 1) [14]. (3) Additionally,
we calculated the straightness of the movement trajectories. As
a movement along the straight line (i.e. along the y-axis) is the
shortest possible trajectory to pass the 10 mm of pulling required
for a successful trial, the absolute area of the error to the straight
line (E) is thought to capture the evolution of movement economy.
(4) Finally, the mean movement velocity (VEL) during the pulling
movement was calculated (Supplementary Fig. 1).
2.5. Statistical analysis
Statistical analyses were performed using Prism version 5.0
(GraphPad Inc., San Diego, CA, USA) and SPSS Statistics 22.0 (IBM
Corp., Armonk, NY, United States). Learning curves for parameters from phase 2 and 3 were compared using 2-way ANOVA for
repeated measurements with group (FP vs. SP) and session (training day 125) as within-subjects factor, as well as the interaction
group session. Post hoc tests were performed using Bonferroni
correction for multiple comparisons. Non-linear tting of all learning curves and time courses shown in this study was performed
using different exponential functions. Functions were selected
based on goodness of t (R2 ) for each particular data set: for
SR y = pl (1 exp (k x)); for VEL: y = y0 + (pl y0) (1 exp
(k x)); for L, VAR, ZCA and E: y = (y0 pl) exp (k x) + pl.
Abbreviations: pl = plateau, k = rise/decay rate, the time constant 
was calculated as the reciprocal of k. All ts presented in this study
had a R2 > 0.5. To quantitatively assess differences between the time
courses of kinematic parameters and/or success rate, rise/decay
rates (k) were compared using a one-way ANOVA. Post hoc tests
were performed using Bonferroni correction for multiple comparisons. Unpaired t-tests were used to compare the number of days
required for absolving phase 2 of training. Numerical results are
expressed as mean and standard error of the mean (SEM).
3. Results
3.1. Conventional measures of performance
Success rate (SR): Animals of both FP and SP groups showed
exponential inter-session improvements in the percentage of successfully executed trials (Fig. 2A). The FP-group shows a steep
learning curve with a low time constant ( FP = 1.22) and a maximal increase in performance between sessions 1 and 2. The curve
plateaus at session 3 and approximates a success rate of 83%
(plFP = 0.83). The learning curve of the SP-group shows a slower
dynamic with a  SP = 2.21 and a maximal increase in performance
between sessions 2 and 3. It plateaus at session 5 and reaches a
comparably lower success rate of 69% (plSP = 0.69). Although effect
of group (F = 21.1; p = 0.0003) as between subject factor is signicant, interaction of group session is not (F = 0.84; p < 0.68). Thus,
there are differences in learning rate and plateau performance

Fig. 1. Experimental setup and behavioral conditions. (A) Schematic drawing of the
experimental setup: the ETH Pattus is positioned in front of a Plexiglas chamber (a).
Animals need to reach through a vertical opening in the front wall (b) to grasp the
manipulandum (c) and interact with the robotic platform. During a movement, the
position of the manipulandum is continuously recorded by high-resolution optical
encoders (d) to assess movement kinematics. In case of a successful movement,
rats get rewarded with a food pellet that is automatically dispensed in a tray at
the back of the Plexiglas chamber (e). At the end of each trial, the end-effector is

retracted and adjusted to the start position by two actuators (f). (B) Representative
photograph of a rat operating the experimental setup. (C) Schematic depiction of the
two behavioral conditions that were tested in this study. In both groups, animals had
to pull the end-effector over a distance of 10 mm along the y-axis. In the free-pulling
group (FP), no restrictions in the x/y-plane were given except for the borders of the
predened workspace of the ETH Pattus (12 mm from the midline). In the straightpulling group (SP), only movements that respected a narrow virtual tunnel (2 mm
from midline, gray shaded) were counted as success. Representative trajectories of
single movements are plotted for the rst (black) and the last (red) training session.
Green dots (top) indicate the start position of the manipulandum in the midline.

100

80

latencies (sec)

O. Lambercy et al. / Behavioural Brain Research 278 (2015) 569576

success rate (%)

572

60
40

SP-group
FP-group

SR

20

20

15
10
5
0

0
0

10

15

20

25

15

20

25

session number

session number

C
100

100

success rate (%)

success rate (%)

10

80
60
40
20

80
60
40
20

session number

session number

Fig. 2. Success and intertrial latencies. (A) Improvement of trial success rate (SR) between sessions was exponential for both groups. FP animals improved faster and reached
a higher plateau-value (signicant effects of group, p = 0.0003). Maximal increase in performance (=maximal learning) occurred between sessions 1 and 2 for the FP-group
and between sessions 2 and 3 for the SP-group. (B) Decay of intertrial latencies (L) is faster and more pronounced in the SP- than in the FP-group (signicant effect of group,
p = 0.0046). (C) To assess intrasession learning, sessions were divided into quintiles according to trial number. The percentage of successful grasps is plotted per quintile. For
SP, a clear within-session improvement can be detected for sessions 13. Interestingly, starting-performance is equal or even increased when compared to the prior session.
For the sake of clarity, error bars were not included in this plot. (D) For FP-group, within-session improvement is only observed in session 2. Session 1 already started with a
success rate of approximately 50%, and shows no evolution over quintiles. For the sake of clarity, error bars were not included in this plot.

between FP- and SP-groups, but the shape of the learning curve is
comparable.
Inter-session grasping latencies (L): L rapidly drops in both groups
(FP-group:  FP = 1.23; SP-group:  SP = 1.14) and approximates low
values (FP-group: plFP = 5.00 s; SP-group: plSP = 3.97 s; Fig. 2B).
ANOVA shows a signicant effect of group (F = 10.67; p = 0.0046)
with a faster and more pronounced drop in the SP-group.
Intra-session improvement: For SP-group (Fig. 2C), a clear
within-session improvement can be detected for sessions 13. Performance levels do not drop below the level of the previous day
and intra-session curves start from an equal or even elevated level
compared with the prior training session. For FP animals (Fig. 2D), a
within-session improvement is only apparent for session 2. In contrast to the SP-group, there is virtually no change in performance for
session 1. In addition, the FP-group already starts at a remarkably
high success rate of approximately 50%.

parameters and SR are both characterized by a low time constant ( FP 1) and a similar progression over time. In contrast,
for SP-group (Fig. 3D), ANOVA analysis of k reveals a highly signicant difference between particular kinematic parameters and
also SR (F = 101.5; p < 0.0001). Post hoc tests even show signicant
differences (p < 0.05) between all parameters except ZCA and success rate. Thus, in the SP-group, different parameters evolve with
distinct temporal dynamics: VAR (SP = 1.48) < ZCA ( SP = 2.62) < E
( SP = 3.39) < VEL ( SP = 5.41). Interestingly, at least for E and VEL,
changes still occur at a time where the classical learning curve
(SR; Fig. 2A;  SP = 2.21) already reached plateau.
4. Discussion
The rodent skilled reaching task (SRT) is frequently used to
study motor learning and motor system plasticity in rats. The animals are trained to reach with their preferred forepaw toward

3.2. Evolution of kinematics

All parameters in both conditions show an exponential
decay/rise with respect to their temporal evolution (Fig. 3A and B).
Learning curves derived from the SP-group are generally characterized by starting points and plateau-values that indicate a higher
degree of skill (in the sense of higher velocity, lower number of zero
crossings, etc.) when compared to the FP-group. Plateau values and
time-constants  characterizing the shape of learning curves are
listed in Table 1.
For the FP-group, the ANOVA analysis of rise/decay rates (k,
as the reciprocal of the time constant ) shows no statistically
signicant difference (F = 1.79; p = 0.157; Fig. 3C). Thus, kinematic

Table 1
Evolution of kinematic parameters. Time-constants and plateau values extrapolated
from exponential rise/decay of the parameters mean variability of trajectories (VAR),
number of zero crossings of acceleration prole (ZCA), error to the straight line (E) and
mean velocity (VEL) of the FP- vs. the SP-group.
Parameter

VAR
ZCA
E
VEL

FP

SP

 FP

plFP

 SP

plSP

1.32
1.27
0.82
1.16

0.20 mm2
8.86
8.79 mm2
1411.6 mm/s

1.15
2.62
3.39
5.41

0.11 mm2
4.64
5.00 mm2
188.5 mm/s

O. Lambercy et al. / Behavioural Brain Research 278 (2015) 569576

VAR

15

20

mean variability in
trajectories (mm2)

mean variability in
trajectories (mm2)

20

SP-group
FP-group

10
5
0
10

15

20

10
5

25

ZCA
20

10

0
5

10

15

20

15
10
5
0
0

10

15

20

150
100

VEL

50

25

20

25

20

25

20

25

0
5

10

15

session number

20

15
10
5
0
0

25

200

20

10

session number

250

15

20

error to straight line (mm2)

20

10

session number

ZCA

25

session number

30

number of zero crossings

30

mean velocity (mm/s)

number of zero crossings

error to straight line (mm2)

session number

mean velocity (mm/s)

VAR

15

0
0

10

15

session number

250
200
150
100

VEL

50

0
0

10

15

20

25

session number

10

15

session number

0.6
0.4
0.2
0.0

SP

VAR
ZCA
E
VEL
success rate

1.0

k (rise/decay rate)

0.8

k (rise/decay rate)

573

FP

0.8
0.6
0.4
0.2
0.0

Fig. 3. Kinematic parameters. (A) and (B) Improvement of kinematic parameters over time is characterized by an exponential rise/decay. For the SP-group learning curves
indicate a higher level of skill (in the sense of higher velocity, lower number of zero crossings, etc.) when compared to FP-group. (C) In contrast for SP, k values vary signicantly
(F = 101.5; p < 0.0001). Post hoc tests show signicant differences (p < 0.05) between all parameters except for ZCA and SR. (D) In the FP-group, the rise/decay rates (k, as the
reciprocal of the time constant ) of kinematic parameters and success rate did not differ (F = 1.79; p = 0.157).

574

O. Lambercy et al. / Behavioural Brain Research 278 (2015) 569576

a food pellet, to grasp and retrieve it [6]. Using this paradigm,

learning-induced changes have been described with respect to
dendritic morphology [15,16], turnover of synaptic spines [17],
gene expression [18,19], synaptic weights [20] and motor cortical maps [21]. To maintain comparability with these studies, key
features of the SRT with respect to workspace, setup and featured
motion sequence (i.e. reachinggraspingretrieval) were maintained while developing the ETH Pattus [10] and the designs of the
motor learning paradigms presented here. Performance gains over
time (learning curve) are commonly measured by the percentage
of successful reaching attempts (i.e. success rate) over a number
of trials. Usually, several training days (sessions) are required to
reach plateau performance and improvement is observed within
and between training sessions [7,12].
In this study, we investigate two reaching paradigms that differ
in complexity: a 10 mm free-pulling task (FP) and a 10 mm straightpulling task (SP) where an additional precision constraint is added
for the movement to be considered successful. Between sessions,
the FP-group quickly gains performance and reaches a high success rate at the plateau (83%). In contrast, the slower time course
of inter-session improvement in the SRT-paradigm [7] with a maximal learning between session 2 and 3 and the approximation of a
plateau at session 5 is comparable to the SP-group presented here.
However, the performance level at plateau is higher in the SP-group
(69%) when compared to SRT (56% [7]). This difference is likely the
consequence of a larger restriction of movements by the ETH Pattus when compared to SRT, assuming that the SRT is more difcult
than SP.
Improvement within training sessions observed for the SRT
[7,22] is also present for session 13 in the SP-group. FP-trained
rats only show intra-session learning in session 2. Interestingly,
for the SP-group, performance levels start from an equal or even
elevated level compared with the performance of the precedent
training session. For SRT, start-performance at day 2 drops below
session 1 end-performance [7]. Thus, at least for the early phase of
learning, consolidation of motor memory seems to be more effective in SP as compared to SRTa further indication that SP is easier
to learn than SRT.
Inter-trial latencies are considered to be a rough measure of
motivation [7]. In all paradigms, latencies quickly drop to plateau
over the rst 13 training days. Although repeated measures
ANOVA indicates a signicant effect of group (SP faster than FP),
Bonferroni-corrected post hoc analysis indicates a signicant difference only for session 1 (p < 0.01). Thus, SP animals operate the
setup signicantly faster, i.e. are more motivated during session 1
when compared to FP animals. This faster handling of the operant
conditioning paradigm embedded in the motor learning task could
be a compensatory strategy to cope with lower success rates and
consecutively rare food reward.
The FP-group differs from the SP-group with respect to the temporal evolution of kinematic parameters: in FP the parameters VAR,
ZCA, E and VEL improve quickly to reach a plateau despite the
fact that SP rats start at better values. Hence, there is less learning occurring in FP as compared to SP. FP may therefore serve as
an activity control for SP, i.e. a task in which the forelimb is used
to a similar amount but this use does not trigger learning to the
same extent. Activity control tasks are often used in motor learning
experiments and are associated with different neuroplastic changes
than learning tasks [16,18,23].
In the SP-group, kinematic parameters develop at different
timescales (Fig. 3C). For humans and non-human primates, several studies support the idea that motor learning is composed of
distinct processes that are represented by different anatomical
substrates and occur over different time frames [24]. Interestingly, no sequential evolution of kinematic parameters is present in
FP-group (Fig. 3D).

In the SP-group, the parameter mean variability in trajectories

(VAR,  SP = 1.48) shows a rapid decay with the smallest timeconstant. Thus, a brief initial phase of trial and error (corresponding
to high variability) is followed by a phase of lower variability. This
process may be interpreted as nding a proper movement strategy
to achieve the goal/reward [13] that is also described for humans
and non-human primates [3].
Number of zero-crossings of the acceleration prole (ZCA,
 SP = 2.62) is a metric of smoothness [14] that provides information about the number of corrective movements [25] and/or
movement-segmentation caused by biomechanical constraints due
to the joint/muscle architecture [26]. With training, the number
of zero-crossings decreases and smoothness improves. This is in
accordance with ndings in humans, showing that the learning of
a task is associated with a progressive blending of a series of discrete
submovements into a compound movement [26,27]. This phenomenon was also observed in developing infants [28] or patients
recovering from brain injury [29]. Thus among all parameters
introduced here the number of zero-crossings may be the best
and most intuitive measure displaying the degree of renement of
a motor program.
The error to the straight line (E,  SP = 3.39) was computed to
reect the economy of movement trajectories. For analytical reasons we assumed that the straight line is the most economic
trajectory, although it may not be the most physiological one [26].
We observed that the straightness of trajectories improves with
learning. Interestingly, this process is characterized by a slower
time constant than inter-session improvement of success rates
( SP = 2.21). Motor systems inherently aim for a state of minimal
energetic cost [8], a principle that also applies for motor learning with the upper extremity in humans [9]. Thus, improvement of
straightness could be an internal mechanism to reduce metabolic
costs by optimizing movement trajectories.
The slowest parameter to improve with learning is movement
velocity (VEL,  SP = 5.41). To maximize movement success, motor
systems can either optimize accuracy (higher probability of a successful trial) or movement speed (higher number of successful
trials by increasing the overall number of trials)the speedaccuracy tradeoff [4]. As motor learning implies increasing both,
accuracy and movement velocity, the speed-accuracy tradeoff
function that accounts for movement precision at dened movement velocities, has been developed [14,30]. Even though this is
a convincing approach, it is difcult to implement into animal
paradigms because of the necessity of varying movement speed
on demand. In primates, precision usually improves before speed
[4,31].
With respect to behavior, the sub-processes of motor learning
described here for rodents share similarities to those described
for humans and primates. However, due to the inter-species differences in organization of motor systems [32,33], it may be
questioned whether they share a similar neural substrate. Our
experimental model is well suited to answer this question, e.g.
by combining behavioral analyses with in vivo electrophysiological measurements and post hoc immunohistochemistry against
cellular markers of learning and activity like cFOS or Arc [18,19].
In addition to the skilled reaching task [6] other paradigms
have been developed to assess learning of skilled forelimb movements in rats [34,35]. Although the utilization of high-speed
cameras allowed the visualization of movement trajectories, velocity proles and degrees of pro/supination [36,37], these endpoints
remained descriptive, not enabling a quantitative analysis. To overcome this limitation, devices to detect or measure certain forelimb
movements have been designed [3840]. However, most of these
devices have limited capacity to sense more complex movements
or actions. Recently, a novel robotic device was used to investigate
recovery in a mouse ischemic stroke model [41]. In this task, rats

O. Lambercy et al. / Behavioural Brain Research 278 (2015) 569576

were required to pull a handle (one-degree-of freedom) over a distance of 10 mm similar to our FP group. The ETH Pattus [10] was
designed particularly to allow more complex forelimb movements
in rats. Due to its three degrees of freedom, movement sequences
can be tested and quantitatively described that are similar to
those required for classical paradigms of motor learning. In addition of measuring movement kinematics, the ETH Pattus is actuated
and can exert forces on the manipulandum while the rat moves it
[42]. This capacity will be tested in future experiments.
Finally, some limitations of the ETH Pattus have to be considered: although the postures an animal can take for interaction with
the robot is limited by the small window through which the animal has to reach and the distance to the end-effector the rat could
approach it with different congurations/positions of its trunk or
forelimb. These elements of motion sequence may also be involved
in the motor learning processes but cannot be properly assessed by
the ETH Pattus in its present form. With respect to this issue, our
robotic platform can complement, but not entirely replace existing
models of motor learning.
5. Conclusion
Using a robotic platform (ETH-Pattus [10,11]), we investigated
motor learning in rats using two forelimb tasks that differ in complexity. The more complex SP condition reects classical paradigms
of motor skill learning in rats such as the skilled reaching task. The
simpler FP condition can serve as an activity control task. With
respect to the temporal pattern and performance level of kinematic
parameters, both conditions can be furthermore reliably differentiated. Analysis of kinematics derived from the SP-group but not the
FP-group furthermore suggest the presence of several independent processes contributing to motor learning. As these processes
share puzzling similarities with concepts gained in humans and
primates, our model is ideally suited to investigate their morphological and physiological bases.
Acknowledgements
This work was supported by the Swiss National Science Foundation (NCCR Neural Plasticity and Repair, Project 310030 135471),
the ETH and the P&K Foundation. The authors are members of the
Neuroscience Center Zurich (ZNZ) and the Rehabilitation Initiative
and Technology Platform Zurich (RITZ). We thank Clement OseiAtiemo for his technical support.
Appendix A. Supplementary data
Supplementary data associated with this article can be found, in
the online version, at http://dx.doi.org/10.1016/j.bbr.2014.10.047.
References
[1] Squire LR. Memory systems of the brain: a brief history and current perspective.
Neurobiol Learn Mem 2004;82(Nov (3)):1717.
[2] Krakauer JW, Mazzoni P. Human sensorimotor learning: adaptation, skill, and
beyond. Curr Opin Neurobiol 2011;21(Aug (4)):63644.
[3] Re GG, Hazen-Martin DJ, El Bahtimi R, Brownlee NA, Willingham MC, Garvin
AJ. Prognostic signicance of Bcl-2 in Wilms tumor and oncogenic potential of
Bcl-X(L) in rare tumor cases. Int J Cancer 1999;84(Apr (2)):192200.
[4] Bogacz R, Wagenmakers EJ, Forstmann BU, Nieuwenhuis S. The neural basis of
the speed-accuracy tradeoff. Trends Neurosci 2010;33(Jan (1)):106.
[5] Hikosaka O, Nakamura K, Sakai K, Nakahara H. Central mechanisms of motor
skill learning. Curr Opin Neurobiol 2002;12(Apr (2)):21722.
[6] Whishaw IQ, Pellis SM. The structure of skilled forelimb reaching in the rat: a
proximally driven movement with a single distal rotatory component. Behav
Brain Res 1990;41(Dec (1)):4959.
[7] Buitrago MM, Ringer T, Schulz JB, Dichgans J, Luft AR. Characterization of motor
skill and instrumental learning time scales in a skilled reaching task in rat.
Behav Brain Res 2004;155(Dec (2)):24956.

575

[8] Todorov E, Jordan MI. Optimal feedback control as a theory of motor coordination. Nat Neurosci 2002;5(Nov (11)):122635.
[9] Huang HJ, Kram R, Ahmed AA. Reduction of metabolic cost during motor learning of arm reaching dynamics. J Neurosci 2012;32(Feb (6)):218290 (the ofcial
journal of the Society for Neuroscience).
[10] Vigaru BC, Lambercy O, Schubring-Giese M, Hosp JA, Schneider M, Osei-Atiemo
C, et al. A robotic platform to assess, guide and perturb rat forelimb movements.
IEEE Trans Neural Syst Rehabil Eng 2013;21(Sep (5)):796805 (a publication of
the IEEE Engineering in Medicine and Biology Society).
[11] Vigaru B, Lambercy O, Graber L, Fluit R, Wespe P, Schubring-Giese M, et al. A
small-scale robotic manipulandum for motor training in stroke rats. IEEE Int
Conf Rehabil Robot 2011;2011:5975349 [proceedings].
[12] Buitrago MM, Schulz JB, Dichgans J, Luft AR. Short and long-term motor skill
learning in an accelerated rotarod training paradigm. Neurobiol Learn Mem
2004;81(May (3)):2116.
[13] Olveczky BP. Motoring ahead with rodents. Curr Opin Neurobiol 2011;21(Aug
(4)):5718.
[14] Shmuelof L, Krakauer JW, Mazzoni P. How is a motor skill learned? Change and
invariance at the levels of task success and trajectory control. J Neurophysiol
2012;108(Jul (2)):57894.
[15] Greenough WT, Larson JR, Withers GS. Effects of unilateral and bilateral training
in a reaching task on dendritic branching of neurons in the rat motorsensory
forelimb cortex. Behav Neural Biol 1985;44(Sep (2)):30114.
[16] Wang L, Conner JM, Rickert J, Tuszynski MH. Structural plasticity within highly
specic neuronal populations identies a unique parcellation of motor learning
in the adult brain. Proc Natl Acad Sci USA 2011;108(Feb (6)):254550.
[17] Xu T, Yu X, Perlik AJ, Tobin WF, Zweig JA, Tennant K, et al. Rapid formation
and selective stabilization of synapses for enduring motor memories. Nature
2009;462(Dec (7275)):9159.
[18] Hosp JA, Mann S, Wegenast-Braun BM, Calhoun ME, Luft AR. Region and taskspecic activation of Arc in primary motor cortex of rats following motor skill
learning. Neuroscience 2013;250(Oct):55764.
[19] Kleim JA, Lussnig E, Schwarz ER, Comery TA, Greenough WT. Synaptogenesis
and Fos expression in the motor cortex of the adult rat after motor skill learning.
J Neurosci 1996;16(Jul (14)):452935 (the ofcial journal of the Society for
Neuroscience).
[20] Rioult-Pedotti MS, Friedman D, Donoghue JP. Learning-induced LTP in neocortex. Science 2000;290(Oct (5491)):5336.
[21] Kleim JA, Barbay S, Nudo RJ. Functional reorganization of the rat motor cortex
following motor skill learning. J Neurophysiol 1998;80(Dec (6)):33215.
[22] Luft AR, Buitrago MM. Stages of motor skill learning. Mol Neurobiol
2005;32(Dec (3)):20516.
[23] Molina-Luna K, Hertler B, Buitrago MM, Luft AR. Motor learning transiently
changes cortical somatotopy. NeuroImage 2008;40(May (4)):174854.
[24] Sakai K, Hikosaka O, Miyauchi S, Takino R, Sasaki Y, Putz B. Transition of brain
activation from frontal to parietal areas in visuomotor sequence learning. J
Neurosci 1998;18(Mar (5)):182740 (the ofcial journal of the Society for Neuroscience).
[25] Milner TE. A model for the generation of movements requiring endpoint precision. Neuroscience 1992;49(Jul (2)):48796.
[26] Flash T, Hogan N. The coordination of arm movements: an experimentally conrmed mathematical model. J Neurosci 1985;5(Jul (7)):1688703 (the ofcial
journal of the Society for Neuroscience).
[27] Burdet E, Milner TE. Quantization of human motions and learning of accurate
movements. Biol Cybern 1998;78(Apr (4)):30718.
[28] von Hofsten C. Structuring of early reaching movements: a longitudinal study.
J Mot Behav 1991;23(Dec (4)):28092.
[29] Rohrer B, Fasoli S, Krebs HI, Volpe B, Frontera WR, Stein J, et al. Submovements
grow larger, fewer, and more blended during stroke recovery. Mot Control
2004;8(Oct (4)):47283.
[30] Reis J, Schambra HM, Cohen LG, Buch ER, Fritsch B, Zarahn E, et al. Noninvasive cortical stimulation enhances motor skill acquisition over multiple
days through an effect on consolidation. Proc Natl Acad Sci USA 2009;106(Feb
(5)):15905.
[31] Hikosaka O, Rand MK, Miyachi S, Miyashita K. Learning of sequential movements in the monkey: process of learning and retention of memory. J
Neurophysiol 1995;74(Oct (4)):165261.
[32] Klein A, Sacrey LA, Whishaw IQ, Dunnett SB. The use of rodent skilled reaching
as a translational model for investigating brain damage and disease. Neurosci
Biobehav Rev 2012;36(Mar (3)):103042.
[33] Nudo R. Neurophysiology of motor skill learning. In: Byrne J, editor. Learning
and memory: a comprhensive reference. Amsterdam: Elsevier; 2008.
[34] Ballermann M, Metz GA, McKenna JE, Klassen F, Whishaw IQ. The pasta
matrix reaching task: a simple test for measuring skilled reaching distance,
direction, and dexterity in rats. J Neurosci Methods 2001;106(Mar (1)):
3945.
[35] Allred RP, Adkins DL, Woodlee MT, Husbands LC, Maldonado MA, Kane JR,
et al. The vermicelli handling test: a simple quantitative measure of dexterous
forepaw function in rats. J Neurosci Methods 2008;170(May (2)):22944.
[36] Whishaw IQ, Pellis SM, Gorny BP, Pellis VC. The impairments in reaching and
the movements of compensation in rats with motor cortex lesions: an endpoint,
videorecording, and movement notation analysis. Behav Brain Res 1991;42(Jan
(1)):7791.
[37] Whishaw IQ. An endpoint, descriptive, and kinematic comparison of skilled
reaching in mice (Mus musculus) with rats (Rattus norvegicus). Behav Brain Res
1996;78(Aug (2)):10111.

576

O. Lambercy et al. / Behavioural Brain Research 278 (2015) 569576

[38] Fowler SC, Davison KH, Stanford JA. Unlike haloperidol, clozapine slows and
dampens rats forelimb force oscillations and decreases force output in a
press-while-licking behavioral task. Psychopharmacology 1994;116(Sep (1)):
1925.
[39] Francis JT, Chapin JK. Force eld apparatus for investigating movement control
in small animals. IEEE Trans Biomed Eng 2004 Jun;51(6):9635.
[40] Duarte JE, Gebrekristos B, Perez S, Rowe JB, Sharp K, Reinkensmeyer DJ. Effort,
performance, and motivation: insights from robot-assisted training of human

golf putting and rat grip strength. IEEE Int Conf Rehabil Robot, 2013 Jun
2013:6650461 [proceedings].
[41] Spalletti C, Lai S, Mainardi M, Panarese A, Ghionzoli A, Alia C, et al. A robotic system for quantitative assessment and poststroke training of forelimb retraction
in mice. Neurorehabil Neural Repair 2014;28(Feb (2)):18896.
[42] Shadmehr R, Mussa-Ivaldi FA. Adaptive representation of dynamics during
learning of a motor task. J Neurosci 1994;14(May (5 Pt 2)):320824 (the ofcial
journal of the Society for Neuroscience).