A Multizone Cerebellar Chip for Bioinspired Adaptive Robot1
Control and Sensorimotor Processing2
Emma D. Wilson, Tareq Assaf, Jonathan M. Rossiter, Paul Dean, John Porrill,
Sean R. Anderson and Martin J. Pearson
3
Abstract4
The cerebellum is a neural structure essential for learning, which is connected via multiple5
zones to many different regions of the brain, and is thought to improve human performance in a6
large range of sensory, motor and even cognitive processing tasks. An intriguing possibility for the7
control of complex robotic systems would be to develop an artificial cerebellar chip with multiple8
zones that could be similarly connected to a variety of subsystems to optimize performance. The9
novel aim of this paper, therefore, is to propose and investigate a multizone cerebellar chip applied10
to a range of tasks in robot adaptive control and sensorimotor processing. The multizone cerebellar11
chip was evaluated using a custom robotic platform consisting of an array of tactile sensors driven12
by dielectric electroactive polymers mounted upon a standard industrial robot arm. The results13
demonstrate that the performance in each task was improved by the concurrent, stable learning14
in each cerebellar zone. This paper, therefore, provides the first empirical demonstration that a15
synthetic, multizone, cerebellar chip could be embodied within existing robotic systems to improve16
performance in a diverse range of tasks, much like the cerebellum in a biological system.17
1 Introduction18
The importance of cerebellar function can be inferred from the estimate that the cerebellum contains19
80% of neurons in the human brain [1]. Interestingly, a single cerebellar microcircuit, with virtually20
identical neuronal circuitry, is repeated throughout the entire cerebellar cortex [2]. Despite this uni-21
formity the cerebellum is implicated in a multitude of diverse tasks. It is traditionally regarded as a22
motor control structure, capable of adaptively modulating motor commands by providing corrections23
required for accurate movements [2, 3, 4]. There is also mounting evidence indicating that the cere-24
bellum is involved in sensory processing, sensory perception and cognitive functions (such as emotion25
and language) [5, 6, 7, 8].26
The microcircuit regularity of cerebellar cortex implies that the same basic signal processing al-27
gorithm is implemented by each region of the cerebellum (a microzone), whether used for control of28
reflexive or voluntary movements, sensory noise cancellation, or higher functions such as language.29
The combination of general applicability with uniformity suggests that functional differences between30
cerebellar microzones emerge from differences in the input and output connectivity [9, 10, 2]. This31
has brought about the ‘chip’ metaphor of cerebellar organisation, where the function of each region32
depends on both the uniform internal algorithm implemented by all chips, and on the architecture33
in which the chip is embedded (e.g. external connections, which differ dependent on function). The34
observation that cerebellar lesions impair but do not abolish function indicates that the cerebellum is35
not the sole pathway subserving each particular function, suggesting that the cerebellum modulates a36
range of behaviours and generally optimises performance [11, 5]. Prior research has also shown that37
serial, or tandem learning enables efficient learning in the cerebellum [12].38
Engineering control solutions are often designed on a case-by-case basis and optimised to a specific39
task. When controlling nonlinear, multi-degree of freedom, compliant, soft robots this is a non-trivial40
problem. Neural control strategies, implemented by structures including the cerebellum, have evolved41
1
over long periods of time to deal with compliant, nonlinear materials such as muscle and biological42
systems are able to achieve remarkable levels of control performance despite using a relatively flat43
and homogenous structure. A synthetic, uniform cerebellar chip algorithm that could be plugged into44
existing systems to fine tune and improve the performance in a range of tasks, much like the cerebellum45
in the biological system, has great potential for robotic applications. Such an algorithm could greatly46
reduce initial control design effort and reduce the need for extensive offline system identification.47
This is because initial control structures would need only to provide an approximate solution (e.g.48
the brainstem in motor plant compensation only approximates the dynamics of the plant) and the49
adaptive cerebellar element could fine tune this control. Such a cerebellar chip would be especially50
useful in lightweight, multi-dimensional, anthropomorphic robots, where considerable control efforts51
are needed.52
Therefore, an intriguing possibility for the control of complex robotic systems with multiple degrees53
of freedom is an artificial ‘cerebellar chip’ that could be plugged into existing control systems to54
fine-tune and improve performance in a wide range of tasks. Previous work in robotics has used55
cerebellar-inspired algorithms to provide adaptive solutions for robot control in single applications56
including variable stiffness, lightweight actuators with varying dynamics [13, 14, 15] especially in the57
context of robot arm control [16, 17, 18, 19, 20, 21, 22]. Cerebellar inspired models have also been58
applied in robotics to locomotion [23], collision, or obstacle avoidance, tasks [19, 24, 25, 26, 27], to59
gaze stabilisation tasks [19, 28, 29], to the adaptive cancellation of self-generated sensory signals [30],60
and to provide anticipatory control [31]. However, cerebellar-inspired control algorithms have not yet61
been tested through simultaneous application of the same microzone algorithm to a range of different62
tasks within a single robotic system. Even in cases when the same model has been applied to different63
sensorimotor tasks in a robotic system [30, 28, 19], this has not been done simultaneously.64
The novel aim of this paper is to propose and investigate simultaneous learning in multiple mi-65
crozones of cerebellar microcircuity applied to different tasks in motor control and sensory processing66
within a robotic system. The impact of interactions between algorithms applied to each distinct task67
is investigated for the first time here. This paper, therefore, represents an essential step towards68
developing a cerebellar chip for robotic systems.69
The cerebellar microzones were each based on the adaptive filter model of the cerebellum [32, 33],70
itself derived from the original Marr-Albus models [3, 4]. This algorithm is computationally powerful71
and is widely used in the analysis of cerebellar function [33]. It can represent both forward and inverse72
dynamic models [9], and has been evaluated in a range of robotic tasks, including image stabilisation73
[28], reafferent noise cancellation [30] and both linear and nonlinear control of artificial muscle [34, 35].74
The following three robot tasks were chosen to investigate: (i) control of an array of artificial75
whiskers using artificial muscle (trajectory control using motor plant compensation); (ii) removal76
of self motion or re-afferent noise from the sensory response of an array of active tactile whisker77
sensors (sensory noise cancellation); (iii) the calibration of a head centred map of sensory space to78
improve accuracy of directed motor commands toward points of interest in that map (sensorimotor79
map calibration). Our design principle, derived from the principle that the same cerebellar algorithm80
is used in a wide variety biological tasks, is that there were to be no ad-hoc changes to the algorithm81
internal circuitry used in each task.82
A custom built robot platform (‘Bellabot’) [36] was used to evaluate the real world performance83
of the cerebellar chips. This platform consists of an array of tactile whisker-like sensors driven by84
artificial muscle actuators (Dielectric ElectroActive Polymers - DEAPs) mounted as the end-effector of85
a standard industrial robot manipulator. DEAPs have inconsistent, time varying, non-linear dynamics86
which present a number of control challenges.87
The remainder of this paper is organised as follows. The following sections describe the robotic88
tasks, platform, and the adaptive filter model of cerebellar function including its application to these89
tasks. The results section provides data on the performance of the cerebellar adaptive filter applied to90
each task concurrently, as well as results showing how concurrent learning affects performance. The91
paper concludes with a discussion of the results.92
93
2
x
-200 -100 0 100 200
y
-200
-100
0
100
200
Figure 1. a) Summary of the basic robotic task. Each of the three areas (motor compensation,
sensory cancellation and map calibration) where cerebellar control is applied are highlighted. The
robotic platform (Bellabot) used to implement the algorithms is pictured, with a picture of it oriented
to look at a target shown on the right of the figure. b) Whisker module activated by DEAPs. The
figures on the left give a top down view of the whisker module and those on the right a side on view.
Each whisker is activated by two separate actuators (shown in black in top down view). Activation of
each side (indicated by on/off label) sequentially drives the whiskers back and forward. Arrows (shown
in the side on view) indicate the direction of movement resulting from the activation of each actuator.
2 Methods94
2.1 Robotic tasks95
The three tasks (trajectory control using motor plant compensation, sensory noise cancellation, sen-96
sorimotor map calibration) were carried out using the robotic platform ‘Bellabot’ (details given in the97
subsequent section). These tasks were chosen as they are distinct, difficult tasks. They are modeled on98
the behaviour of whiskered rodents attending to points of salient contact made by their whiskers [37].99
Challenges include, non-linearities and time variations in the motor plant and reafferent response, and100
mapping of points of whisker contacts to the true head centred topographic location.101
The Bellabot platform used an array of artificial whiskers, mounted upon a manipulator, to detect102
novel tactile targets. The location of the detected targets was represented in a topographic map of103
whisker sensory space and then used to drive orienting movements towards the target. Accuracy of the104
orienting movements was assessed using a camera located at the centre of the whisker array. Orienting105
errors were used to update learning in a cerebellar module calibrating the topographic map. Additional106
cerebellar modules were used to control the trajectory of each whisker (modeled on rodent whisking107
behaviour [38]) and to remove re-afferent, self-induced noise signals from each whisker during target108
detection. A picture of the robot platform and summary of the three basic tasks in the context of109
detecting and orienting towards novel objects is given in Fig. 1a and the individual whisker module in110
Fig. 1b. To evaluate the performance of each adaptive filter over time data were obtained while the111
Bellabot cyclically performed four sequential behaviours: explore, recoil, orient and reset. The full112
protocol duration was 40 minutes, during which these four cyclic behaviours were repeated.113
Before each trial a small ball (target), mounted upon a clamp stand, was placed in front of the114
platform (see left hand picture in Fig. 1a). This was placed approximately 200-300mm out, within115
a radius of 100mm from the centre of the camera. During the explore behaviour (average duration116
4.5 secs/ cycle), the array of whiskers were actively driven with the noise cancelling and trajectory117
tracking chips active and learning.118
The manipulator was moved forward until a contact was made by one of the whiskers (if no119
3
contact was made the Bellabot was reset and the target re-placed). During the contact (i.e. when120
the deflection signal was above threshold) learning was gated, and the adaptive filter weights were not121
updated. Detected targets were read from the 2D topographic map of whisker sensory space.122
After contact the recoil behaviour (average duration 8.3 secs/ cycle) was initiated, where the123
robot moved backwards a safe distance from the target. During the recoil behaviour active whisking124
continued, with the noise cancelling and trajectory tracking chips active and learning (to increase the125
time period for learning C
m
and C
s
weights - Sections 2.4, 2.5).126
The head was then moved to orient (average duration 5.9 secs/ cycle) such that the centre mounted127
camera was directed towards the estimated target point in space as determined from the head centred128
topographic map of the whisker sensory space.129
At the end of the orienting movement an image of the ball was captured from the camera. A130
coloured blue ball (that differed from the colours in the background) was used as a target. The centre131
of the target was then calculated from the image as the centre of mass of the blue area of pixels. This132
estimated target location was used to obtain the target acquisition error and update the weights of the133
map-calibration adaptive filters. During image capture the whiskers were held stationary and the noise134
cancelling and trajectory tracking adaptive filters were not active, and filter weights not updated. The135
Bellabot was then reset and the platform returned to its original start configuration before beginning136
another trial.137
Between trials the contact ball was relocated pseudo-randomly such that different whiskers were138
contacted in subsequent cycles, with a similar number of total contacts being made on each whisker.139
To simplify the experiments, we only considered cases where the inner circle of eight whiskers were140
contacted (contacting just the inner eight whiskers was decided as a trade-off between testing the141
map calibration algorithms works and limiting the length of experiments to keep manageable). The142
behavioural cycle was repeated 85 times to obtain a rich set of data to calibrate the topographic map.143
2.2 Robotic Platform144
A custom built robotic platform was used to evaluate the real world performance of each task. The145
Bellabot (Fig. 1a) platform consisted of an array of 20 DEAP actuated whiskers surrounding a central146
camera, mounted on a 5 degree-of-freedom industrial manipulator (ABB - IRB120) (for further details147
see [39, 36]).148
Each DEAP driven whisker was made of a single conical membrane with two carbon grease elec-149
trodes printed onto the membrane. The whisker actuation was constrained to 1-dof, and whisker150
deflections and rotations were measured using Tri-axis Hall effect sensors. Signals from these sensors151
were sampled at 500Hz and passed via USB to an external computer. Motor commands to move152
each whisker were relayed from this computer and converted to the high voltages required for DEAP153
actuation.154
A standard USB camera mounted at the centre of the whisker array was used to capture a series155
of images at the end of an orienting movement to calculate errors in the x−, y− directions, i.e. the156
difference between the centre of the image (desired position) and the centre of the object actually157
oriented to (actual position).158
The coordination of motor command and sensory response to and from the platform was main-159
tained using the BRAHMS modular execution framework [40]. BRAHMS was developed specifically160
to integrate heterogeneous models of neural components by providing standardised interfaces and161
maintaining synchronous communication and execution across these models suitable for application in162
hard-real-time constrained robot control. The core of BRAHMS was written and compiled using C++,163
however, it has numerous language wrappers for users to adopt including, python, C# and MATLAB,164
and can be deployed using MPI for parallel compute environments. In this study the cerebellar micro-165
circuits were coded using C++ and the interface to robot control modules in C#, offline data analysis166
was performed using MATLAB. Raw sensory data gathered from the whisker array was sampled and167
marshaled at 500Hz before being downsampled to 25Hz for presentation to the modularlized algorithms168
under test, the coordination for which is one of the native operations of BRAHMS.169
4
2.3 Cerebellar Microcircuit170
mossy fiber
parallel fiber
PF/PC synapse
climbing fiber
Purkinje cell
granule
cell
C
0
C
Figure 2. The cerebellar microcircuit as an adaptive filter. (a) Simplified cerebellar microcircuit. (b)
Interpretation of the cerebellar microcircuit as an adaptive filter. Mossy fibre inputs (u) are analysed
into component parallel fibre signals (g
i
) in the granule layer by a bank of fixed filters. We use a bank
of fixed alpha basis filters (see Eq. 6). These parallel fibre signals are weighted (w
i
) and recombined to
give the Purkinje cell output (z). (c) The cerebellar microcircuit as an adaptive filter with additional
Q matrix applied to filter outputs (g
i
) to approximately orthonormalise into parallel fibre signals (p
i
).
In the basic cerebellar microcircuit (Fig. 2a) there are two input pathways (climbing fibres and171
mossy fibres) and a single output (Purkinje cells). In the adaptive filter model of cerebellar function172
(Fig. 2b), mossy fibre inputs are analysed into component parallel fibre signals which are weighted173
(parallel fibre - Purkinje cell synapses) and recombined to form the Purkinje cell output. The climbing174
fibre inputs are a teaching or error signal, used to train the weights of parallel fibre-Purkinje cell175
synapses [32, 33]. An extension to the adaptive filter model (Fig. 2c) used here, includes a fixed176
matrix Q to orthonormalise
1
signals and speed up learning.177
In interpreting the cerebellar microcircuit as an adaptive filter, the time-varying signals carried178
by mossy fibre inputs to the cerebellum are represented at sample time T as u(T ) (for clarity only179
a single input is considered here, i.e. the number of mossy fibre inputs n
u
= 1). These inputs are180
passed through a basis of n fixed filters G
i
to produce signals g
1
(T ), g
2
(T ), ..., g
n
p
(T ) (where n
p
is the181
number of parallel fibre signals, in this single mossy fibre input case n
p
= n, for cases where there are182
more mossy fibre inputs n
p
= n × n
u
). A fixed matrix Q is used to orthonormalise signals to produce183
1
This orthonormlisation step is described in further detail in [35]. Although necessary to speed up learning, the use
of a fixed matrix is not a biologically plausible mechanism and future work is required to establish how biology solves
the problem of fast learning
5
