This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2018.2799145, IEEE Access

Versatile locomotion control of a hexapod robot

using a hierarchical network of non-linear oscillator
circuits
Ludovico Minati Mattia Frasca Natsue Yoshimura, Yasuharu Koike
(Senior Member, IEEE) (Senior Member, IEEE) FIRST
Tokyo Tech World Research Hub Initiative, Dept. of Electrical Institute of Innovative Research
Institute of Innovative Research Electronic and Tokyo Institute of Technology
Tokyo Institute of Technology Computer Engineering (DIEEI) Yokohama 226-8503, Japan
Yokohama 226-8503, Japan University of Catania
and Complex Systems Theory Dept. 95125 Catania, Italy
Institute of Nuclear Physics
Polish Academy of Sciences
Kraków 31-342, Poland
Email: [email protected],
[email protected], [email protected]

Abstract—A novel hierarchical network based on coupled non- tion gait, Non-linear oscillator, Synchronization.
linear oscillators is proposed for motor pattern generation in
hexapod robots. Its architecture consists of a Central Pattern I. I NTRODUCTION
Generator (CPG), producing the global leg coordination pattern,
coupled with six Local Pattern Generators (LPGs), each devoted A. Walking robots
to generating the trajectory of one leg. Every node comprises a Walking robots are of interest from both application and
simple non-linear oscillator and is well-suited for implementation research perspectives: on one hand, they offer unsurpassed
in a standard Field-Programmable Analog Array (FPAA) device.
resilience and maneuverability over irregular terrains, reaching
The network enables versatile locomotion control based on
five high-level parameters which determine the inter-oscillator places inaccessible to rovers, and on the other they yield
coupling pattern via simple rules. The controller was realized important insights into animal behavior and neurophysiology,
on dedicated hardware, deployed to control an ant-like hexapod since many species possess four or more legs, and adaptive
robot, and multi-sensory telemetry was performed. As a function motor control is arguably among the most important func-
of a single parameter, it was able to stably reproduce the
tions supporting survival [1]–[4]. Yet, their viability remains
canonical gaits observed in six-legged insects, namely the wave,
tetrapod and tripod gaits. A second parameter enabled driving constrained by the complexity inherent in controlling their
the robot in ant-like and cockroach-like postures. Three further locomotion, a task that has stimulated and continues to stim-
parameters enabled inhibiting and resuming walking, steering, ulate research aiming to develop circuit-based as opposed to
and producing uncoordinated movement. Emergent phenomena purely algorithmic controllers. The purpose of efforts in this
were observed in the form of a multitude of intermediate gaits,
area is not only to replicate electronically the biological neural
and of hysteresis and metastability close to a point of gait
transition. The primary contributions of this work reside in architectures which solve the locomotion control problem for
the hierarchical controller architecture and associated approach a variety of structures across diverse scales (e.g., from micro-
for collapsing a large set of low-level parameters, stemming insects to large mammals), but also to practically realize
from the complex hexapod kinematics, into only five high-level high levels of flexibility and adaptability (e.g. responding to
parameters. Such parameters can be changed dynamically, an
unpredictable stimuli, failures etc.) which can only be attained
aspect of broad practical relevance opening new avenues for
driving hexapod robots via afferent signals from other circuits with recourse to emergent phenomena, about the induction and
representing higher brain areas, or by means of suitable brain- management of which still little is known [5]–[8].
computer interfaces. An additional contribution is the detailed
characterization via telemetry of the physical robot, involving B. Artificial central pattern generators
the definition of parameters which may aid future comparison Analog circuit-based approaches to walking robot control
with other controllers. The present results renew interest into are often inspired by the highly influential neurobiological
analog CPG architectures and reinforce the generality of the
connectionist approach. discovery of Central Pattern Generators (CPGs). These are
Index Terms—Analog locomotion control, Bio-inspired control, pervasive architectures which intrinsically generate periodic
Central Pattern Generator (CPG), Field-Programmable Analog oscillations, and drive the leg joints via other downstream
Array (FPAA), Hexapod robot, Hierarchical network, Locomo- neural structures. In most legged animals, CPGs are capable

2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2018.2799145, IEEE Access

of creating a multitude of stereotyped gaits, i.e. regular phase lowest cost but also largest size and lowest flexibility [19],
relationships between leg movements, such as walking and [20]. Much work has since been done on implementations
trotting, adapted to support diverse essential behaviors such involving ad-hoc designed CMOS ICs, substantially reducing
as foraging and fleeing over highly heterogeneous surfaces. size and opening the way to increased flexibility through
As consistently indicated by lesion studies, the generation of allowing seamless reconfiguration of connections and alter-
rhythmic patterns occurs spontaneously without need for any ation of oscillator parameters via pass-gates and switched-
external input; however, biological CPGs invariably receive capacitor circuits [13], [27]. Custom CMOS ICs have enabled,
afferences from other ganglia or higher brain regions, which for example, the compact implementation of spiking neural
influence motion initiation and inhibition, and determine the models [15] as well as of a multitude of non-linear oscillators
gait that is generated alongside its frequency and other pa- and cellular neural networks [4], [21]. Such devices are,
rameters. In particular, CPG activity can often be externally however, still relatively inflexible, because it is difficult to exert
modulated to produce slow drifts in gait or sudden, discrete control over large numbers of parameters without developing
movements, though the exact mechanisms underlying the latter prohibitively complex configuration logic. This represents an
remain incompletely understood [9]–[12]. important limitation, since for research as well as practical
A rich literature on implementing artificial CPGs in the form applications one often requires a high level of flexibility
of electronic circuits is available [4], [8], [12], [13]. While in controlling the global pattern generation and even the
reviewing it is beyond the scope of the present study, we operation of individual legs. Furthermore, the availability of
shall mention that three broad approaches based on non-linear custom ICs is by definition limited to the centers where they
dynamical systems have been considered: 1) circuits pred- are developed and collaborating institutions, hampering the
icated upon physiologically-realistic models of spiking and replication of studies.
bursting neurons [14], [15]; 2) circuits encompassing lower- Recently, some Field Programmable Gate Array (FPGA)-
order coupled oscillators with isolated periodic dynamics, based implementations of CPG circuits have also been re-
which although biologically less plausible often yield a more ported; although notable in terms of size and power, concep-
compact representation having similar emergent properties tually they are equivalent to numerical simulations on stan-
[4], [16]–[21]; 3) circuits involving chaos control, wherein dard computers and as such share the same limitations [28].
interaction with the environment stabilizes the oscillation of By contrast, Field-Programmable Analog Arrays (FPAAs)
one or more chaotic oscillators into a periodic orbit, and as a are switched-capacitor systems which architecturally resemble
function of one or more parameters such orbit can be altered programmable logic devices (PLDs), but implement analog
to yield distinct gaits [22], [23]. rather than digital circuitry, and enable the rapid physical
We note that in circuits of coupled oscillators the gait proper- deployment of near-arbitrary circuits based on only few
ties can oftentimes be influenced via two separate routes: 1) commercially-available, general-purpose arrays. Representa-
acting on one or more internal parameters of each oscillator tive devices include a multitude of arrays of interconnected and
(i.e., of the Hopf, Matsuoka, van der Pol or Kuramoto oscil- configurable analog blocks, each of which comprises 1) a set
lators as in Refs. [16]–[18]), or 2) changing the topology and of operational amplifiers, 2) a set of capacitors having diverse
strengths of the connections coupling the oscillators, such as in values, 3) switch matrices allowing the realization of almost
Refs. [4], [19]–[21]. We refer to the latter as the connectionist any combination of connections between the amplifiers and the
approach. capacitors, 4) one or more banks of static memory adapted
Besides investigations in which the robot and environment to store a bit-string representing the current configuration
were both simulated, also a considerable fraction of experi- and 5) means for externally uploading new configurations
mental studies, wherein the robot was physically realized, have without disrupting operation [29]–[32]. Countless non-linear
resorted to numerical implementation of CPG circuits, as this oscillators have been realized using this technology (see Ref.
minimizes the engineering challenges associated with building [33] for a review), and a recent study has introduced a circuit
the circuit and adjusting its structure and parameters to obtain conjugating a rich dynamical repertoire with compactness [26].
the desired behaviors. However, there is inherent value in However, to the authors’ knowledge an FPAA-based CPG
implementing CPGs by means of analog electronic circuits: implementation is missing to date.
not only this usually results in much smaller and lower-power
systems, but there are also a multitude of aspects which C. Scope and structure of this study
can have a profound impact on global emergent properties In this work, a versatile controller for a hexapod robot
in systems of synchronized non-linear elements. These, in was developed and experimentally realized in the form of
particular, include the stability of numerical solutions, the a hierarchical network of coupled oscillators controlled by
effects of discretization and noise, as well as non-ideal be- five high-level parameters P1 . . . P5 . As detailed below, in
haviors, which are not trivial to capture numerically, and the this initial study these parameters are kept static, i.e. are set
parametric mismatches making the constituent elements non- once before each experimental run, but in the future they will
identical [20], [24]–[26]. be controlled dynamically, i.e. as P1 (t) . . . P5 (t), for example
Early physical realizations of CPG circuits were based on off- based on circuits representing higher brain areas or input from
the-shelf components such as operational amplifiers, yielding brain-computer interfaces.

2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2018.2799145, IEEE Access

A preliminary evaluation of the controller was performed postures, it is possible to collapse them into as few as five
based on a physically-realized hexapod robot, and a range high-level parameters, which specify in a continuous manner
of kinematic measurements are presented to illustrate its which configurations are active at a given time. Fourth, we
capabilities. Suitability for efficient implementation in FPAAs report on a comprehensive experimental characterization in-
was a key factor driving the design of the proposed controller, volving telemetry of a physical robot, and the calculation of
with the aim of overcoming many of the above-mentioned measures which also warrant future consideration as means
limitations associated to realizations based on custom ICs (i.e., for comparing controllers in this area.
availability and reconfigurability) or FPGAs (i.e., digital vs. The remainder of this paper is organized as follows. In
analog implementation). Though the proposed FPAA-based Sec. II the controller network is introduced and its design
CPG realization could in principle be generalized to driving and dynamics are illustrated. In Sec. III the hexapod robot
diverse robot types (e.g. bipeds, quadrupeds, even swimming design and the experimental results obtained driving it with
robots), here the focus is restricted to developing a specific the proposed controller are described. Finally, in Sec. IV the
controller for hexapod robots. main features of the proposed approach are discussed with
The approach is predominantly a connectionist one, wherein reference to other known controllers.
according to simple mapping laws, the five high-level control
II. C ONTROLLER N ETWORK
parameters P1 . . . P5 influence the strengths of a large number
of links (realized as input gains), realizing a considerable level A. Architecture
of flexibility in terms of gait pattern, level of coordination, 1) Hierarchical network: As mentioned in the previous
posture and steering control. A hierarchical approach, inspired section, the proposed controller consists of two hierarchical
by the architecture of biological CPGs, was applied to realize a levels: at the top, one CPG, and at the bottom, six LPG struc-
control system having at its core a network of six oscillators, tures, one per leg (Fig. 1). The CPG operates spontaneously,
each one corresponding to a leg of the robot and coupled in other words it does not receive any external afferences and
through downstream unidirectional connections to a separate is instead controlled by five high-level parameters. Since the
(i.e., local) pattern generator individually driving the joints of focus in this initial study was on motor pattern generation,
the associated leg. the parameters are set once and kept static (i.e., constant)
The topology of the circuit instanced at each node can be during each experimental run. As illustrated in greater detail
viewed as a ring oscillator onto which two integrators are over- below, they control gait selection (P1 ), walking initiation and
laid; saturation at the outputs of latter represents the main type inhibition (P2 ), strength of coupling between the CPG and
of non-linearity in the system. In a previous study, coupling the LPGs (P3 ), posture (P4 ) and steering (P5 ); a summary is
such oscillators revealed non-trivial phenomena including a provided in Table I.
form of “remote synchronization”, wherein amplitude fluc- The CPG provides the overall pattern, i.e. determines the phase
tuations become entrained between non structurally-adjacent relationships between the leg swing and stand cycles. The CPG
nodes, leading to the emergence of small-world synchroniza- is unidirectionally coupled to the LPGs, which are independent
tion topology from underlying structural connectivity arranged of each other and receive no other input. They provide a
as an elementary ring network [26]. The possibility of easily flexible means of translating the signal generated by the CPG
building coupled configurations encompassing more than one into three signals driving the joints of each leg, determining
unit and of exploring large regions in the multidimensional the precise trajectory of the coxa-body (α), tibia-femur (β)
space of the parameters of the system, is fundamental to the and femur-coxa (γ) joints during each cycle. Each LPG is
present work: it provides the substrate for controller versatility, internally wired as a ring, allowing it to potentially generate
which is practically realized, as mentioned above, by mapping “intrinsic” sustained activity irrespective of “extrinsic” input
five high-level control parameters onto a large number of gains from the associated CPG node, which is unsynchronized
(connection strengths) across the two layers of the hierarchical between legs due to unavoidable parametric mismatches (i.e.,
controller. tolerances) between the physical oscillators (Fig. 1).
This work aims to provide the following contributions with The proposed controller is specifically intended for driving
respect to the existing literature. First, we address the issues hexapod robots with 18 degrees-of-freedom, a representative
of flexibility and availability associated with CPGs deployed in implementation of which is introduced in Sec. III for the
custom ICs, and demonstrate the possibility for seamlessly re- sole purpose of experimentally confirming the viability of the
configurable implementation in FPAAs. Second, we introduce controller. Throughout this paper, the six legs of a generic
a pure hierarchical approach where robot motion is controlled hexapod robot are denoted with L1, . . . , R3, where L/R
by a CPG which downstream drives independent local pattern stands for left/right, the leg pairs are numbered 1-3 from front
generators (LPGs) instanced for each leg, and demonstrate to rear, and each leg has three degrees of freedom whose
that this approach confers a high level of flexibility in gait, corresponding angles are referred to as α, β and γ.
posture and coordination. Third, we demonstrate a means Under the control of the proposed network, a hexapod robot
of dealing with the large number of low-level parameters can be made to walk in “ant-like” and “cockroach-like”
inherent in specifying one CPG and six independent LPGs: postures; in the former, the power stroke is delivered using
by superimposing a minimal number of canonical gaits and the coxa-body joint of all legs, whereas in the latter, the

2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2018.2799145, IEEE Access

Fig. 1. Hierarchical controller design. Overall network architecture, delineating a central pattern generator (CPG, 6 nodes) and six local pattern generators
(LPGs, 3 nodes each). Each leg is associated to one CPG node (e.g., node L1/C) driving a corresponding LPG (e.g., nodes L1/α, L1/β and L1/γ), which
in turn drives a leg (note: figure layout not corresponding to physical leg layout). The link strengths among the CPG nodes depend on the gait selection
parameter P1 and on the activation parameter P2 , and the link strengths within each LPG depend on the coupling strength parameter P3 and on the posture
parameter P4 . Different colors highlight which links have increased intensity for a given gait (CPG) or configuration (LPGs); the full relationships between
parameters and connections are specified in Tables II, A.I and A.III.

0.8
2) Central pattern generator (CPG): The architecture of
0.6 the CPG is predicated on the work of Arena and colleagues
0.4 [4], [34], wherein the wave, tetrapod (metachronal) and tripod
0.2
gaits were implemented in custom CMOS ICs applying the
connectionist approach, that is by means of altering the
0 coupling between oscillators while maintaining their internal
0 0.2 0.4 0.6 0.8 1
parameters unaltered. The coupling schemes realizing these
three gaits are taken directly from their work, however flipping
Fig. 2. Membership functions for the wave gait SL (P1 ), tetrapod or link directions to account for the fact that the oscillator circuit
metachronal gait SM (P1 ), and tripod gait SH (P1 ) as a function of gait
selection parameter P1 chosen for implementation in this study, described below, has
predominantly integrative dynamics and as such yields a phase
lead rather than a lag or a delay.
The wave gait is implemented through six unidirectional
front and hind legs deliver the power stroke primarily using positive connections having strength SL and linking the fol-
the tibia-femur joint. Regardless of the posture, locomotion lowing loop of nodes L1/C → R3/C → L2/C → R1/C →
can be attained by means of several gaits, including three L3/C → R2/C → L1/C (green arrows in Fig. 1). The
“canonical” gaits. With this term, we refer to gaits which tetrapod (metachronal) gait is implemented through eight
are highly conserved in insects, namely 1) the tripod gait, unidirectional positive connections having strength SM or
wherein two leg triplets, i.e. (L1,R2,L3) and (R1,L2,R3), SM /2 and linking the following two intertwined loops of
conjointly alternate in their stance and swing phases, 2) the nodes: L1/C → L2/C → L3/C → R2/C → L1/C and
wave gait, characterized by the power stroke sequence L1, R1/C → R2/C → R3/C → L2/C → R1/C (blue arrows in Fig.
R2, L3, R1, L2, R3, and 3) the tetrapod or metachronal gait, 1). The tripod gait is implemented through five bidirectional
wherein the power stroke sequence is (L1,R3), R2, (L3,R1), negative connections having strength −SH or −SH /2 and
L2, i.e., the legs L1 and R3 and the legs L3 and R1 swing linking the following ladder of nodes: L3/C ↔ R1/C ↔
simultaneously. In insects, the wave, tetrapod and tripod gaits R2/C ↔ L2/C ↔ L1/C ↔ R3/C (red arrows in Fig. 1).
tend to be observed in conjunction with slow, medium and fast The successful generation of the intended gait patterns based
locomotion respectively [1], [2]. on the coupling schemes introduced in Refs. [4], [34] despite

2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2018.2799145, IEEE Access

TABLE I
S UMMARY OF THE HIGH - LEVEL CONTROL PARAMETERS P1 . . . P5 . I N THIS INITIAL STUDY THEY ARE SET ONCE BEFORE EACH EXPERIMENTAL RUN
AND KEPT STATIC THEREAFTER , BUT IN FUTURE WORK THESE CAN BE CHANGED DYNAMICALLY AS P1 (t) . . . P5 (t) TO DRIVE THE ROBOT
PURPOSEFULLY. CPG: C ENTRAL PATTERN G ENERATOR . LPG S : L OCAL PATTERN G ENERATORS .

Parameter Range Target Denomination Function

P1 [0, 1] CPG Gait selection parameter Determines the phase relationships between legs, e.g. allows choosing between wave
(slow), tetrapod (metachronal, medium speed) and tripod (fast) gaits
P2 [−1, 1] CPG Activation parameter Allows inhibiting and approximately reversing the CPG activity, i.e. walking
P3 [0, 1] LPG CPG
LPG coupling Determines the level of synchronization between the CPG and the LPGs (i.e., movement
strength parameter coordination)
P4 [0, 1] LPG Posture parameter Allows choosing between the ant-like and cockroach-like postures
P5 [−1, 1] LPG Steering parameter Allows steering the robot trajectory sideways

profound differences in the oscillator circuit instanced at each

node reinforces the generality of connectionist approach.
Extending the notion of continuous generalized gait introduced
in Refs. [4], [34], the CPG connection strengths are made to
depend upon a single gait selection parameter P1 ∈ [0, 1].
Namely, instead of having two separate mixing parameters,
here a single high-level parameter akin to a speed setting,
namely P1 , univocally sets the link strengths between all CPG
nodes in such a manner as to realize the three canonical gaits.
This is accomplished with recourse to the three sigmoidal
membership functions SL (P1 ), SM (P1 ) and SH (P1 ), which
reach unity respectively for P1 ≈ 0, P1 ≈ 1/2 and P1 ≈ 1:

1
 SL (P1 ) = 1 − 1 + eAL (P1 +CL )





 1
SM (P1 ) = 1 − AM (|P1 +CM |+CM /2)
(1)

 1 + e

 1
 SH (P1 ) =


1 + eAH (P1 +CH )
where empirically we set AL = AM = AH = −20, CL = −0.3
and CM = CH = −0.5; these settings, which are not critical,
yield membership functions maximizing the responsiveness
of the controller to parameter P1 (Fig. 2), as ascertained
in preliminary experiments not reported for brevity. In this
manner, for P1 ≈ 0, P1 ≈ 1/2 and P1 ≈ 1 the canonical
gaits are recovered, whereas other settings allow continuous
transition between them via a multitude of “intermediate”
gaits.
As shown in Fig. 1, each CPG node is hardwired to receive
input from up to three other CPG nodes, and linear superpo-
sition of the connections required for the three canonical gaits
yields the continuous generalized gait. As detailed below, this
is realized by acting on three parameters separately across all
CPG nodes, namely gains G1 , G2 and G3 .
The oscillator instanced at each CPG node and implemented
in a corresponding FPAA device represents an adaptation
of that described in Ref. [26]. Namely, the circuit consists
of i) two gain and low-pass filtering stages which mix the
inputs from other CPG nodes, yielding internal voltages v1
and v2 , ii) a ring of three similar gain and low-pass filtering
Fig. 3. Non-linear oscillator structures. a) Oscillator circuit instanced at each stages, yielding internal voltages v3 , v4 and v5 , and iii) two
CPG node (Eqn. (2)). b) First circuit variant, instanced at the LPG α nodes integrators overlapped to the ring, yielding internal voltages
(Eqn. (4)). c) Second circuit variant, instanced at the LPG β and γ nodes v6 and v7 (Fig. 3a). In particular, each oscillator receives as
(Eqn. (5)). See text for detailed descriptions.
input the voltages v5 of three other oscillators, referred to as

2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2018.2799145, IEEE Access

vA , vB and vC , which are weighted by the input gains G1 , the original ring connectivity, the filter frequencies and
G2 and G3 , respectively, mixed with the node’s own output integration constants were gradually scaled, in small steps,
and fed into the ring as internal voltage v2 .1 The available to bring oscillation from the low-kHz down to the low-Hz
gait patterns are determined by the hardwired interconnections spectral region, 2) at each step, the gains were manually
among inputs and outputs of the CPG nodes, and the specific adjusted to maintain an approximately constant signal
gait generated at a given time is determined by the associated amplitude at each internal node, 3) after reaching the target
strengths (gains). The interconnection scheme realizing the frequency range, the connectivity was changed to that in Fig.
graph in Fig. 1, alongside the corresponding gains, is specified 1 and the parameters were further tuned, one by one, to obtain
in Table II. the desired synchronization behavior in the CPG and LPG
The system of ordinary differential equations corresponding structures. Steps 1) and 2) were completed in simulations,
to each CPG node can be written as follows: since access to all internal nodes was necessary, whereas step

dv1   3) was completed deploying the oscillators to the physical

 = Γ 2πF 1 (G 1 vA + G 2 vB + G 3 vC − v1 ), v1 circuit board, since consideration of the realistic behavior
dt





 dv2   was necessary. The process was labor-intensive and relied

 = Γ 2πF 2 (G 4 v1 + G 5 v 5 − v2 ), v2 on trial-and-error largely driven by intuitive appreciation of
dt




 dv3   circuit and network behavior, representing a known issue
−


 = Γ 2πF 3 (G v
6 2 v3 ), v 3 associated to controller of this kind: at present there is still
 dt



dv4   very limited knowledge about how to induce and manage
= Γ 2πF4 (G7 v3 + G8 v6 − v4 ), v4 (2) desired emergent phenomena [5]–[8].
 dt


 dv5   The only tunable parameters in the CPG nodes are the input
= Γ 2πF5 (G9 v4 + G10 v7 − v5 ), v5




 dt gains G1 , G2 and G3 , which are set according to Table II,
and the internal gain G4 = −P2 , where P2 ∈ [−1, 1] is the


 dv6  
= Γ K v , v

1 2 6



 dt activation parameter, which provides control over whether
to enable the CPG by setting P2 ≈ 1, disable it by setting

 dv7
  

 = Γ K2 v3 , v7 P2 ≈ 0 or generate an approximately backwards gait pattern
dt
where for better readability we have omitted the subscript by setting P2 ≈ −1.
referencing the CPG node label, i.e., v1 = v1,L1/C , v2 = v2,L1/C , Since they are based on switched-capacitor circuits, FPAAs,
etc. for node L1/C associated to leg L1, and so on. We note albeit analog, are discrete-time devices, therefore in numerical
that the only non-linearity appearing in Eqn. (2) is the function simulations Eqn. (2) was accordingly discretized according to
Γ(x, y), which represents saturation due to finite voltage swing Euler’s method, setting ∆τ = 6.25 ms as the fixed step size,
according to corresponding to a switching rate of fc = 160 Hz.
Two additional remarks on node dynamics are required. First,
Γ (x, y) = R (x) H(Vs − y) − R (−x) H(Vs + y) (3) while minimizing the number of state variables is always
desirable, given a predetermined oscillator circuit such as
where H(x) = 1 for x > 0, 0 for x ≤ 0 and R(x) = xH(x). the present one the number of equations is inherently set
Following empirical considerations, to realize self-sustained by the circuit topology. In this case, the internal signals are
oscillation in each CPG node at a frequencies suitable for clearly partially correlated, making the “effective” number
driving the physical robot in real-time, the filter frequencies of variables lower than the number of physical variables.
are set to F1 = 15 Hz, F2 = F3 = F4 = F5 = 0.5 Hz, the Attempts to simplify the numerical model of the circuit are
integration constants are set to K1 = 5.5 × 10−6 µs−1 and beyond the scope of this work, and from an implementation
K2 = 1.3 × 10−6 µs−1 , the gains are set to G5 = −0.5, perspective we note that by design one circuit instance fits
G6 = 1.2, G7 = 0.8, G8 = 0.6, G9 = 1.5, G10 = −0.4, exactly into a single FPAA. Second, as discussed in Ref.
and with the intended target FPAA devices, the maximum [26] and in Sec. III, not all saturation nonlinearities present
voltage at any node is |Vs | = 4 V. Due to manufacturing in Eqn. (2) are indispensable to support the generation of
tolerances, all parameters are subject to a random variation periodic orbits and stable gait patterns, those at the integrator
on the order of 0.1%-1% between nodes. Compared to the outputs clearly prevailing over all others. Nevertheless,
initial work wherein this circuit was introduced [26], here given that in the physical device saturation can occur for any
the parameter values and the coupling scheme are different, variable, here we retained the Γ(x, y) function in all equations.
yielding oscillation at considerably lower frequencies and
stable periodic orbits over a wide range of settings. 3) Local pattern generators (LPGs): The architecture of
Given the presence of such large number of parameters and the LPGs is inspired by the observation that in nature some
unavailability of a formal synthesis method, we proceeded as situations and types of neurological damage (e.g., focal lesions
follows, starting from the circuit in Ref. [26]: 1) maintaining involving motor circuits, blockade of certain receptor types)
1 We highlight that in the present implementation the factor of −0.8
can produce uncoordinated leg movement, demonstrating the
introduced by the output inverter is transparently taken into account when action of “local” circuits which can generate sustained activity
configuring the FPAAs to realize the prescribed coupling gains. even when deprived of the afferences from the CPG, and

2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2018.2799145, IEEE Access

TABLE II
H ARDWIRED CPG NODE CONNECTIVITY AND CORRESPONDING WEIGHTS ( I . E ., INPUT GAINS ), CONTROLLED BY PARAMETER P1 , WHICH IS A
HIGH - LEVEL SETTING AS SPECIFIED IN TABLE I. VOLTAGES vA , vB AND vC DENOTE THE THREE INPUTS OF EACH NODE ( SEE F IG . 1), WEIGHTED
RESPECTIVELY BY THE GAINS G1 , G2 AND G3 ( SEE F IG . 3), WHEREAS VOLTAGES v5 CORRESPOND TO THE NODE OUTPUTS . F OR BREVITY, SH , SM AND
SL STAND FOR SH (P1 ), SM (P1 ) AND SL (P1 ) ACCORDING TO E QN . (1). CPG: C ENTRAL PATTERN G ENERATOR .

For node vA vB vC G1 G2 G3
L1/C v5,R3/C v5,L2/C v5,R2/C −SH /2 −SH /2 SM + SL
R2/C v5,L2/C v5,R1/C v5,L3/C −SH /2 (SM − SH )/2 SM /2 + SL
L3/C v5,L2/C v5,R1/C v5,R3/C SM SL − SH 0
R1/C v5,L2/C v5,R2/C v5,L3/C SM + SL −SH /2 −SH /2
L2/C v5,L1/C v5,R3/C v5,R2/C (SM − SH )/2 SM /2 + SL −SH /2
R3/C v5,L3/C v5,R2/C v5,L1/C 0 SM SL − SH

at the same time is driven by the aim of allowing flexible necessary to realize all desired phase relationships between
implementation of the available gaits through more than one the joints.
power stroke delivery scheme, or “posture” [1], [2]. As shown For the LPGα oscillators (Fig. 3b), the resulting system of
in Fig. 1 and previously introduced, each LPG comprises ordinary differential equations can be written as follows:
three nodes, denoted with α, β and γ, whose outputs provide 
dv1  
the position set-points for the servo-motors actuating the 
 = Γ 2πF 1 (G 1 vL + G 2 vC + G 3 v5 − v1 ), v1
dt


corresponding coxa-body, tibia-femur and femur-coxa joints 


 dv2  
of the associated leg. 
 = Γ 2πF 2 (G 4 V r + G5 v 5 − v2 ), v2
dt


The strength of coupling between each CPG node and the 

 dv  
3
= Γ 2πF3 (G6 v1 − v3 ), v3

associated downstream LPG nodes is set by the coupling 

 dt


strength parameter P3 ∈ [0, 1]; as demonstrated below, this 
dv4  
parameter influences the level of synchronization between the = Γ 2πF4 (G7 v3 + G8 v6 − v4 ), v4 (4)
 dt
outputs of the CPG and the corresponding downstream LPGs. 

 dv5  
= Γ 2πF5 (G9 v4 + G10 v7 − v5 ), v5

When P3 ≈ 1, the temporal variance in the signals generated 


 dt
by each LPG is almost entirely determined by the CPG input 

 dv 6
 
= Γ K v , v

(extrinsic activity, supported by the connections represented 


 dt
1 1 6
as yellow arrows in Fig. 1), whereas when P3 ≈ 0, the LPG 
 dv  
 7 = Γ K2 v3 , v7


nodes oscillate spontaneously and asynchronously with respect dt
to CPG activity, therefore without stable phase relationships
between the legs, leading to a “thrashing-like” uncoordinated where F2 = 15 Hz, G9 = −1.5, G10 = 0.4, Vr = 3 V and all
movement (intrinsic activity, supported by the connections other parameters are set as for the CPG nodes.
represented as purple arrows in Fig. 1). For the LPGβ and LPGγ oscillators (Fig. 3c), the resulting
The LPGs are also controlled via the posture parameter system of ordinary differential equations can be written as
P4 ∈ [0, 1], which enables transitioning between the ant-like follows:

and cockroach-like postures: for P4 ≈ 0 (ant-like posture), dv1  

 = Γ 2πF 1 (G1 vL + G 2 v C + G 3 v5 − v1 ), v1
the power stroke is delivered by the coxa-body joint of all dt




legs, whereas for P4 ≈ 1 (cockroach-like posture), the front

 dv2  

 = Γ 2πF 2 (G4 v1 + G 5 v5 − v2 ), v 2
and hind legs deliver the power stroke primarily using the dt



dv3
  
tibia-femur joint, in opposite directions. Finally, the steering

= Γ 2πF3 (G6 v1 − v3 ), v3



parameter P5 ∈ [−1, 1] enables steering the trajectory of the  dt



robot sideways by reducing the power stroke amplitude on dv4  
= Γ 2πF4 (G7 v3 + G8 v6 − v4 ), v4 (5)
either side. 
 dt
dv5
  
The parameters P3 , P4 and P5 , together with a set of constants

= Γ 2πF5 (G9 v4 + G10 v7 − v5 ), v5



specified below, influence the dynamics of the LPG nodes


 dt
through the gains G1 , G2 , . . . , G5 in the LPG circuits, shown

 dv6  
= Γ K1 v1 , v6


dt

in Figs. 3b and c. These circuits present minor structural



differences with respect to the CPG oscillators; in particular,
 dv 
 7 = Γ K2 v3 , v7




the LPGα nodes, controlling the coxa-body joints, have the dt

capability of adding a large offset Vr to the output, which without further changes to the parameters.
is necessary to make the front and rear legs “spread out” In all LPG nodes, the gains G1 , G2 and G3 are determined by
with respect to the body, whereas the LPGβ and LPGγ nodes, parameters P3 and P4 according to the B parameters defined
controlling the other two joints, have the capability of mixing by the expressions in Table A.I: G2 controls the coupling
the output of the ring v5 directly with the summation of the of the LPGs to the CPG, whereas G1 and G3 respectively
external inputs v1 , which is referred to as “mixed output” and control the interdependence between the nodes of each LPG

2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2018.2799145, IEEE Access

and the internal loop gain within each node; variable vL corresponding LPG direct outputs, 3) the β joint servo-motors
corresponds to the output voltage of the previous node in the of the front and hind legs are driven by the corresponding
LPG ring (i.e., voltage −v5 for α
β coupling, −4v2 for LPG direct outputs, those of the middle legs are driven by
β
γ and γ
α coupling), and variable vC corresponds to the corresponding LPG mixed outputs, 4) the γ joint servo-
the v5 voltage of the CPG node associated to the LPG (see motors of all legs are driven by the corresponding LPG mixed
Fig. 1). outputs. Third, the LPG
servo-motor connectivity required
The expressions in Table A.I represent the linear superposition for uncoordinated movement (P3 = 0) is implemented in
of three distinct canonical configurations, controlled by the same manner as walking in the ant-like posture, but
different B parameter subsets. To aid the understanding of with different parameter settings. In addition, steering (P5 )
these parameters, individual descriptions are provided in is implemented exclusively by reducing power stroke width,
Table A.II. First, the phase relationships required for walking on the α or β joints for ant- and cockroach-like walking
in the ant-like posture (P3 = 1, P4 = 0) are implemented respectively.
as follows: 1) the LPG α and β nodes of all legs receive
the afferent signal from the associated CPG node, 2) the 4) Parameter determination: To drive the experimental
LPG γ nodes of all legs instead receive the output of the robot described in the next section, the constants in Table A.I
corresponding β node, and 3) the LPG α and γ nodes of and A.III are empirically set as follows: B1 = 1.2, B2 = 1.5,
all legs have non-zero internal loop gain. Second, the phase B3 = B5 = −B9 = 0.8, B4 = −2, B6 = −2.2, B7 = 3,
relationships required for walking in the cockroach-like B8 = 2, C1 = −0.28, C2 = −C3 = C12 = −C13 =
posture (P3 = P4 = 1) are implemented as follows: 1) the C20 = −C21 = 0.04, C4 = C5 = −C6 = C7 =
LPG α and β nodes of all legs and the γ nodes of the front C14 = −C15 = C22 = C23 = −C24 = C25 = 0.1,
and hind legs (opposite phase) receive the afferent signal C8 = C16 = C26 = 0.36, C9 = C17 = C19 = C27 = 0.3
from the associated CPG node, 2) the LPG γ nodes of the and C10 = −C11 = −C18 = C28 = −C29 = 0.2.
middle legs instead receive the output of the corresponding β The procedures for determining these parameters are described
node, and 3) the LPG α and γ nodes have non-zero internal in the Appendix. It should be noted that these settings plausi-
loop gain for the middle legs, whereas the LPG β nodes bly do not represent neither a unique nor an optimal solution.
have non-zero internal loop gain for the front and hind legs. They were not subject to any optimization or robustness as-
Third, uncoordinated movement (P3 = 0) is implemented as sessment; they simply reflect a configuration of the controller
follows: 1) there is no input from the CPG, 2) the LPG nodes which is viable for the specific robot under consideration. In
are coupled forming the following loop α
β
γ
α, and future deployments of the controller, these parameter settings
3) the LPG α nodes of all legs have non-zero internal loop may be tuned using techniques such as genetic algorithms, as
gain. described in Ref. [18], and even the parameter formulations
Similarly, the gains G4 and G5 are determined by parameters can likely be improved. While targeting different mechanics
P3 , P4 and P5 according to the C parameters defined by the could require extensive re-determination of the C parameters,
expressions in Table A.III; together, G4 and G5 influence how the B parameter settings plausibly represent a starting point
the servo-motor position signals, corresponding to voltage v2 , of good general validity.
are obtained from the internal voltages v1 and v5 . To simplify
the notation in Tables A.I and A.III, we have set P30 = 1 − P3 , B. Dynamics
P40 = 1 − P4 , P50 = 1 − R(−P5 ) and P500 = 1 − R(P5 ). 1) Initial simulations and physical implementation: To
The expressions in Table A.III also represent the linear have access to all internal voltages v1 , . . . , v7 of each node,
superposition of three distinct canonical configurations, the CPG subnetwork was initially simulated according to
controlled by different C parameter subsets. To aid the Eqns. (1)-(3) using a fixed-step solver and the parameter
understanding of these parameters, individual descriptions settings prescribed above. Generation of the expected gait
are provided in Table A.IV. First, the LPG
servo-motor patterns at low, medium and high settings of P1 was observed.
connectivity required for walking in the ant-like posture However the wave gait was initially obscured by onset of
(P3 = 1, P4 = 0) is implemented as follows: 1) the α joint global synchronization [24], which was manifest as in-phase
servo-motors of all legs have no offset (coxa-body angle movement of all legs (data not shown). This could be avoided
zero-point is such that the legs do not “spread out”), 2) the α by having random vi (0) ∈ [−Vs , Vs ] and adding a Gaussian
joint servo-motors of all legs are driven by the corresponding random variation in the CPG parameters having a standard
LPG direct outputs, 3) the β and γ joint servo-motors of all deviation of 0.5%, a situation which recalls observations
legs are driven by the corresponding LPG mixed outputs. from the initial study on this oscillator, wherein parametric
Second, the LPG
servo-motor connectivity required for mismatches were fundamental for the emergence of structured
walking in the cockroach-like posture (P3 = P4 = 1) is synchronization patterns instead of global synchronization
implemented as follows: 1) the α joint servo-motors of the [26]. Consideration of the individual voltage time-series,
front and hind legs are held in fixed positions with negative exemplified in Figs. 4a and b, indicated that saturation to
and positive offsets respectively (these legs “spread out”), 2) ±Vs appeared primarily at the integrator outputs v6 and
the α joint servo-motors of the middle legs are driven by the v7 (and, for low settings of P1 , also at v5 ), which is in

2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2018.2799145, IEEE Access

agreement with the observations reported in Ref. [26], despite The ability of the CPG to respond to abrupt parameter
the widely different parameter settings. Indeed, the presence changes was subsequently tested by dynamically altering
of non-linearity primarily or solely in the form of saturation P1 and P2 in large steps. The results, exemplified in Figs.
is a recurring feature across many of the known FPAA-based 4d and e, demonstrated that the CPG was able to rapidly
non-linear oscillators [33]. switch between gaits without instability, as well as to inhibit
The entire network shown in Fig. 1 was thereafter physically and resume oscillation reliably. However, while the gait
implemented on the LYAPUNOV-1 circuit board, which changes shown in Fig. 4d were near-immediate, the build-up
effectively constitutes a reconfigurable analog machine in the of oscillations starting from an “isoelectric” condition was
form of a plug-in card for a standard desktop computer; it gradual, requiring a time-interval equivalent to several cycles.
provides 32 FPAAs alongside supporting infrastructure for As better illustrated in Figs. 7a and 8a respectively for
real-time data acquisition and dynamic reconfiguration of all P1 = 0.8 and P1 = 0.2, the tripod gait emerged slowly
analog circuit parameters. The design is publicly available through the build-up of initially disordered oscillations which
and described in detail in Refs. [26], [35]; here, the FPAA eventually developed the desired phase relationships, whereas
interconnections on the circuit board were manually rewired the wave gait emerged more rapidly, through symmetry
as pictured in Fig. 5a, and detailed in the online materials breaking caused by one node which, after a certain delay,
[36]. Compared to numerical simulations, deployment on generated a large fluctuation and triggered the wave. As shown
physical hardware knowingly results in richer dynamics, for in Fig. 4e, sudden transition to an approximately reverse gait
example due to parametric mismatches, non-ideal behavior of induced by setting P2 = −1 caused a perturbation persisting
circuit elements and noise; in particular, physically Γ(x, y) for a time interval similar to that required for initial oscillation
does not correspond to an ideal step function due to more build-up. Even though these phenomena are in themselves
complex circuit behavior as saturation is approached [26], interesting, they could be viewed as a shortcoming when
[33]. rapid controllability is the purpose; in such cases, the role of
Since FPAAs are switched-capacitor circuits, and hence feedback through sensory signals from peripheral receptors
have discrete-time dynamics, it is possible to effectively becomes important for speeding up convergence towards the
“rescale” the temporal dynamics simply by changing the desired steady-state solution. Injecting a suitable transient
clock frequency fc = 1/∆τ within the range allowed by the in the oscillator parameters, e.g. with P2 > 1, could also
hardware, representing a notable advantage in comparison to represent a viable approach to implement faster start-up.
circuits built with discrete components [4], [20], [33]; in this
work, all experiments not involving real-time control of the 3) Transitions and synchronization: To gain further insight
physical robot were performed with a time-step reduced by a into the transitions between the gaits, the continuous phase of
factor of 5 to limit the acquisition time. each leg ϕi (t) was extracted by calculating, for each of the
six CPG outputs i ∈ L1/C, R2/C, . . . , R3/C, the corresponding
2) Gait pattern generation: To test gait pattern generation analytic signal as follows
on the physical oscillator network, the parameter P1 was first
v5,i (t) + iv̂5,i (t) = Ai (t)eiϕi (t) (6)
continuously swept in a cycle between 0 and 1, dynamically
changing in small steps the gains G1 , G2 and G3 according to where v̂5,i is the Hilbert transform of v5,i (t)
Table II. As shown in Fig. 4c, the CPG responded gradually, 1
Z ∞
v5,i (τ )

generating initially the wave gait for P1 < 0.25, then the v̂5,i (t) = p.v. dτ (7)
π −∞ t − τ
tetrapod (metachronal) gait for 0.25 < P1 < 0.45, then a
“paradoxical” intermediate gait featuring inversion of the and where p.v. denotes the Cauchy principal value of the
phase relationships between legs L3-R1 and R3-L1 for integral [37].
0.45 < P1 < 0.65, then disordered activity for P1 ≈ 0.65 and The characterize each gait, predicated on Wilson’s rules [1],
finally the tripod gait for P1 > 0.75; a hysteresis effect was [2], the following summary parameter was introduced
also observed, as described in detail below. Corresponding ∆ϕ1 + ∆ϕ2 − ∆ϕ3 + ∆ϕ4 + ∆ϕ5 − ∆ϕ6
ζ= P (8)
gait patterns observed when starting and operating the CPG ∆ϕ
at constant P1 settings are shown in Figs. 6a-g and were where
largely similar, albeit with some deviation in the unstable 
∆ϕ1 = |hϕL1/C (t) − ϕR2/C (t)i∆t |
region around P1 ≈ 0.65. These results confirm that the CPG


= |hϕR2/C (t) − ϕL3/C (t)i∆t |

is capable of expressing a generalized gait as a function of



∆ϕ 2
= |hϕL3/C (t) − ϕR1/C (t)i∆t |

∆ϕ
P1 , featuring not only the hardwired canonical gaits, but also 3 (9)
additional emergent intermediate gaits which, as demonstrated ∆ϕ4
 = |hϕR1/C (t) − ϕL2/C (t)i∆t |
in the following section, were kinematically viable. This

= |hϕL2/C (t) − ϕR3/C (t)i∆t |


∆ϕ5
result is particularly noteworthy, as it confirms the ability of


= |hϕR3/C (t) − ϕL1/C (t)i∆t |

the controller to generate behaviors beyond the hardwired ∆ϕ6
ones. The raw time-series recorded at all network nodes are and where hi denotes the temporal average over a sufficiently
publicly available [36]. long observation time, i.e. ∆t  ∆τ . As shown in Fig. 4f,

2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2018.2799145, IEEE Access

ζ > 0 for the wave, tetrapod (metachronal) and intermediate downstream LPG α, β and γ nodes, and averaged across all
gaits, peaking at ζ ≈ 1 for P1 ≈ 0.4, whereas ζ < 0 for six legs according to
the tripod gait. Consideration of this parameter revealed clear
hysteresis in the transition between the slower gaits and the hrC→L i = (rL1/C,L1/α + rL1/C,L1/β + rL1/C,L1/γ +
tripod gait, which occurred in the region of 0.6 < P1 < 0.7: rR2/C,R2/α + rR2/C,R2/β + rR2/C,R2/γ +
depending on the sweep direction, the summary parameter rL3/C,L3/α + rL3/C,L3/β + rL3/C,L3/γ +
ζ flipped sign at different points, namely P1 ≈ 0.63 and
P1 ≈ 0.67, and the disturbance accompanying the transition rR1/C,R1/α + rR1/C,R1/β + rR1/C,R1/γ +
manifested with different phase relationships between the legs. rL2/C,L2/α + rL2/C,L2/β + rL2/C,L2/γ +
In contrast with that between the wave and the tetrapod rR3/C,R3/α + rR3/C,R3/β + rR3/C,R3/γ )/18 (12)
(metachronal) gait, this transition was therefore not smooth,
plausibly because it involved “competition” between links As shown in Fig. 4i, this parameter responded to P1 in a
having positive (SL , SM ) and negative (SH ) weights, as per Ta- manner similar to hrC i, largely irrespective of the posture
ble II. In addition to hysteresis, metastability, that is presence setting (i.e., P4 = 0 vs. P4 = 1). As shown in Fig. 4j,
of states with a finite life-time, was observed, as exemplified the same gradually increased with the coupling strength P3 ,
for P1 ≈ 0.66 in Fig. 4h. The transition was therefore which set the unidirectional energy transfer rate maintaining
unequivocally a first-order one. These phenomena, which were synchronization between the CPG and the LPGs; for P3 > 0.3,
not explicitly designed for but spontaneously emerged from the complete synchronization was consistently observed, with
connectivity and dynamics, are particularly striking as they hrC→L i ≈ 1, whereas at lower P3 settings two different
mirror experimental observations on the transition between critical coupling strengths could be observed, corresponding
the same gaits in living insects and mathematical models to P3 ≈ 0.1 for the wave gait and P3 ≈ 0.3 for the tripod gait
predicated on completely different formulations (for example, [24].
see Ref. [38]). Altogether, these results confirm i) the synchronizability of the
Similarly, as shown in Fig. 4g, the emergent period of the gaits hierarchical oscillator network across all gaits, and ii) the abil-
τ , defined on the basis on the first peak of the autocorrelation ity to gradually transition the LPGs between a configuration
function, was lowest for the tripod gait, which in nature is in which they generate “intrinsic” activity unsynchronized to
generally associated with fast walking [2]; as expected, τ the CPG (low P3 ) and one in which they closely track the
diverged in the region of the hysteretic transition. “extrinsic” input from the CPG (high P3 ).
We subsequently considered the level of synchronization, The raw time-series recorded for all network nodes are pub-
expressed in terms of phase-locking according to licly available [36].

ri,j = |hei[ϕi (t)−ϕj (t)] i| (10) III. E XPERIMENTAL D EPLOYMENT

A. Robot Design and Data Acquisition
where ri,j ∈ [0, 1] (the extreme values representing, respec-
To demonstrate the viability of the proposed controller
tively, complete asynchrony or presence of a perfect phase
experimentally, the same was deployed to perform real-time
relationship, regardless of the relative phase angle and of cycle
open-loop control of an ant-like hexapod robot, named 非
amplitude, between two oscillators i and j [39]).
線形蟻-1 (hisenkei ari-1), derived from the A-Pod design by
First, this parameter was calculated and averaged between all
Kåre Halvorsen (a.k.a. “Zenta”; marketed by Lynxmotion Inc.,
CPG node pairs (voltages v5 in Fig. 3a), according to
Swanton VT). The robot, shown in Fig. 5b, was equipped with
six identical legs, each providing three degrees of freedom
hrC i = (rL1/C,R2/C + rL1/C,L3/C + rL1/C,R1/C +
corresponding, as discussed above, to the coxa-body (α), tibia-
rL1/C,L2/C + rL1/C,R3/C + rR2/C,L3/C + rR2/C,R1/C + femur (β) and femur-coxa (γ) joints. The robot included
rR2/C,L2/C + rR2/C,R3/C + rL3/C,R1/C + rL3/C,L2/C + additional joints for the mandibles and tail, which were not
rL3/C,R3/C + rR1/C,L2/C + rR1/C,R3/C + rL2/C,R3/C )/15 (11) used in the present experiment and were set to constant angles.
Each axis was actuated by an HS-645MG servo-motor (Hitec
which indicated that oscillations remained strongly coherent RCD Inc., Chungcheongbuk-do, South Korea), providing a
among the CPG nodes, i.e. hrC i ≈ 1, for all settings of maximum stall torque of 9.6 kg · cm. The robot was 65 cm
P1 except in proximity of the hysteretic transition, where, as long and 45 cm wide, and weighted 3.25 kg; the tibia and
discussed above, activity became disordered and the period τ femur were respectively 14 cm and 10.5 cm long. The center-
accordingly diverged (data not shown). of-mass was ≈ 2 cm forward of the middle legs, and the body
Second, phase variables for the LPG nodes were introduced raised 9 cm above the base plate. The robot was powered by
in a way similar to Eqn. (6), considering the voltage v2 a 5000 mAh, 6.0 V nickel–metal hydride battery pack.
for nodes L1/α, L2/α, . . . , R3/α (Fig. 3b) in order to define The LYAPUNOV-1 circuit board (Fig. 5a) was installed in
ϕL1/α , . . . , ϕR3/α , and so on for the β and γ nodes (Fig. a desktop computer running Linux, and the experiment was
3c). The phase-locking index as in Eqn. (10) was thereafter programmed using MATLAB and C code. A first thread man-
calculated between each CPG node and the three associated aged FPAA array configuration and real-time data acquisition,

2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2018.2799145, IEEE Access

Fig. 4. Central pattern generator (CPG) dynamics. a) and b) Simulated time-series from all internal voltages of a representative CPG node, respectively
for gait selection parameter settings P1 = 0.8 and P1 = 0.2. c) Gait diagram, representing experimental CPG output time-series, for a continuous cycle
P1 = 0 → 1 → 0 in 0.1 steps; black and white areas respectively denote the swing and stance phases; data length for each parameter setting step ∆t ≈ 13 s.
d) and e) Gait diagrams, respectively for sharp transitions of the gait selection parameter P1 = 0.2 → 0.8 → 0.2 → 0.4 and of the activation parameter
P2 = 1 → 0 → 1 → −1; data length for each parameter setting step ∆t ≈ 60 s. f) and g) Gait summary parameter ζ and period τ as a function of P1 ;
blue and red: distinct identical runs. A complete cycle P1 = 0 → 1 → 0 is charted, and vertical arrows denote the direction of discontinuous transition for
ζ. h) Gait diagram exemplifying spontaneous transition observed for P1 = 0.66. i) Average phase synchronization between the CPG and corresponding LPG
nodes hrC→L i as a function of P1 ; blue and red: ant-like and cockroach-like postures, i.e. posture parameter P4 = 0 and P4 = 1, respectively. A complete
cycle P1 = 0 → 1 → 0 is charted. j) Same parameter hrC→L i as a function of the coupling strength parameter P3 ; blue and red: wave and tripod gaits,
i.e. P1 = 0.2 and P1 = 0.8, respectively, solid and dashed: ant-like and cockroach-like postures, i.e. P4 = 0 and P4 = 1, respectively. A complete cycle
P3 = 0 → 1 → 0 is charted. Unless otherwise specified P2 = P3 = 1. Corresponding data files available online [36].

2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2018.2799145, IEEE Access

Fig. 5. Experimental setup. a) Semi-transparent view of the LYAPUNOV-1 board, depicting the allocation of network nodes to FPAAs, and the re-wiring
implementing the CPG and LPG links (blue, yellow). b) Hexapod robot, featuring independent coxa-body (α), femur-coxa (γ) and tibia-femur (β) joints
for each leg. Locations of the strain gages for measuring the femur longitudinal strains εi are indicated by “*”. CPG: Central Pattern Generator. FPAA:
Field-Programmable Analog Array. LPG: Local Pattern Generator.

and forwarded via TCP/IP the FPAA output measurements was synchronized with the acquisition of telemetry data from
(digitized at 12-bit, 160 Hz) to a second thread, which per- all other sensors. The robot was operated over a smooth
formed down-sampling from 160 Hz to 16 Hz, linear mapping floor in an indoor environment where additional landmarks
to the individual servo-motor angles (calibration), enforced had been installed; as discussed below, this represents an
joint angle limits and streamed the signals to the robot via important limitation of this initial study, in that the open-loop
a dedicated radio link. A third thread gathered telemetry and configuration did not provide any adaptiveness which would
video data. The robot was locally controlled by a network allow locomotion over irregular terrain. The elevation of the
of five ATmega328 micro-controllers (Microchip Technology body over the ground was not derived from the video data,
Inc., San Jose CA), programmed in the Arduino environ- and instead it was directly measured with higher precision
ment (Arduino S.r.l., Ivrea TO, Italy), which implemented by means of a time-of-flight laser sensor installed under the
motor control, sensor interfacing and radio communication bottom plate (VL53L0X; STMicroelectronics S.p.A., Agrate
functions. The robot was interfaced to the computer by Brianza MI, Italy).
means of an isochronous frame-based protocol updating at 16 The video frames were de-multiplexed, de-interlaced, con-
frames/second and deployed over separate up- and down-link verted to gray-scale and undistorted, after which feature identi-
channels (APC220; AppCon Technologies, Shenzhen, China). fication and matching was performed by means of the speeded-
All source code is publicly available [36]. up robust features (SURF) technique [41]. Camera rotation
Pitch and roll data were gathered from an inertial motion unit and translation were thereafter estimated using the maximum
(IMU) based on direction cosine matrix (DCM) formation likelihood estimation sample consensus (MLESAC) approach,
(9 DoF Razor IMU; SparkFun Inc., Niwot CO) [40]. To derived from the RANSAC estimator [42]. Following initial
perform visual odometry, stereoscopic images were acquired evaluation, it was decided to perform monocular odometry
from a camera (BlackBird 2; FPV3DCAM, Moscow, Russia) separately for the two cameras, combining the respective pro-
mounted above the mandibles, having focal length 2.5 mm, cessed data for each frame, iteratively choosing the time-span
stereo base 42 mm and resolution 680 × 512 per eye, which for motion estimation, and applying median filtering to the six

2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2018.2799145, IEEE Access

resulting series of Euler angles and displacements. To improve (data not shown). This result confirmed that the outputs of
accuracy, distance calibration was performed separately for the hierarchical controller have a high level of periodicity,
each experimental run, taking as reference a forward-facing i.e. gait generation is largely stable. It was however noted
ultrasonic sensor mounted under the mandibles (MB1242; that oscillation start-up for P1 = 0.6 occasionally fails,
MaxBotix Inc., Brainerd MN). Because accuracy validation highlighting a form of instability in the transition region.
was not performed, the odometry data should nevertheless be An “idealized” walking robot reacts immediately to the con-
considered as merely illustrative. troller outputs and has no memory, thus when driven by
To gain further insight into the stability of the gaits, each femur perfectly regular signals, all of its physical variables change in
was instrumented with a strain gage (MMF003247; Micro- an identical manner between each gait cycle and the next one.
Measurements Inc., Wendell NC) sensitive to longitudinal However, a physical robot such as the present one has a certain
strain, and connected to an operational amplifier installed weight distribution, finite joint torques, elasticity and a number
proximally. The resulting time-series εi (t) provided an indi- of non-idealities, meaning that it may not be able to reproduce
cation of the level of strain in the plastic structure of the legs, faithfully the controller outputs, for example if the gait period
induced by both acceleration and contact with the terrain. As τ is too short, which may lead to structural instability evident
specified in the next subsection, the focus was on the temporal in the form of fluctuations between one gait cycle and the
regularity, hence no calibration was performed and the values next. To investigate this aspect, the parameter η was calculated
were treated as arbitrary units. from the femur strain signals εi (t) recorded in all gaits
and postures. For this and the subsequent analyses, segments
B. Analyses and Results of undisturbed walking (i.e., excluding the initial transient,
Initial experiments were performed suspending the robot contact with obstacles etc.) were manually segmented from the
atop an elevated platform, which allowed optimal observation experimental data and concatenated. As indicated in Table III,
of the phase relationships between the legs; as documented the observed values were overall high, i.e. η ≈ 0.9, suggesting
in Fig. 6 and associated video materials [36], the actuators that the robot could successfully reproduce the prescribed
successfully reproduced the gait patterns generated for P1 ∈ gait patterns, however some intermediate gaits and the tripod
[0.2, 0.8], P2 = P3 = 1, P4 = P5 = 0. Subsequently, the robot gait were relatively unstable in the cockroach-like posture
was exercised in 22 experimental runs on indoor environment, (i.e., runs n = 15 . . . 18). As expected, during uncoordinated
detailed in Table III, which spanned the same range of P1 movement (i.e., runs n = 21, 22) the leg strain signals were
separately in the ant-like and cockroach-like postures, i.e. temporally disordered, i.e. η ≈ 0.5.
P4 = 0 and 1 respectively, and additionally covered the cases Representative gait patterns, leg joint position signals, tra-
of steering, i.e. P5 = −1, and uncoordinated movement, i.e. jectories and video frame sequences for walking in the ant-
P3 = 0 (see Table I for parameter definitions). like and cockroach-like postures are shown, respectively, in
In the absence of external perturbation of the control param- Figs. 7a-e and Figs. 8a-e. The leg trajectories in the space
eters, if the controller is stable one expects the generation of of the [α, β, γ] joint angles, visible in Fig. 7c and Fig. 8c,
a perfectly regular (i.e., periodic) gait pattern. An empirical delineated clear convergence to different limit cycles after
approach to testing this hypothesis is by means of quantifying initial transient stabilization. By contrast, the trajectory for
how much variance in the generated time-series si (t) can be the case of uncoordinated movement, visible in Fig. 9c had
expressed as a partial Fourier series up to a given finite number a significantly more complex structure, which was unrelated
of terms k. Such periodicity ratio may be written as follows to CPG output and resembled a strange attractor; however,
convincing signatures of self-similarity could not be detected
η = hσ 2 [ŝi (t)]/σ 2 [si (t)]ii∈L1,R2,...,R3 (13) upon calculation of the local slopes of the correlation sum (cor-
where σ 2 [s(t)] denotes the variance of s(t) for 0 < t < tmax , relation dimension analysis as in Ref. [26]; data not shown).
and the corresponding partial Fourier series The femur strain signals εi (t) corresponding to these cases,
shown in Fig. 7f, Fig. 8f and Fig. 9f, revealed markedly lower
k k
a0 X X regularity under the condition of uncoordinated movement
ŝ(t) = + an cos(nt) + bn sin(nt) (14) compared to the other two; it should be noted that, within
2 n=1 n=1
each experimental run, the time-course differences between
where, in the present case, we set k = 8 and an and bn legs reflected multiple factors such as the weight distribution
were estimated by robust non-linear least-squares optimization on the robot and the controller parameter settings.
over multiple runs. The periodicity ratio η can be considered Further emergent features of the gaits could be garnered by
as an order parameter, since one expects η ≈ 0 and 1, considering the pitch θ(t), roll φ(t) and the body elevation
respectively, for stochastic and ordered dynamics. Away from over terrain ẑ(t) (where ˆ denotes world frame coordinate).
the transition region and given sufficient time for initial Considering the variances of pitch and roll, i.e. σ 2 [θ(t)] and
transient stabilization, i.e. for large tmax , for all CPG and LPG σ 2 [φ(t)], as measures of kinematic stability, it was observed
outputs we observed η ≈ 1, even though slow oscillations that walking in the ant-like posture was generally more stable,
extending beyond the gait period occasionally emerged at particularly for the tripod gait (e.g., Fig. 7g), and primarily
some LPG outputs, plausibly due to parametric mismatches associated with pitch oscillation, whereas walking in the

2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2018.2799145, IEEE Access

TABLE III
S UMMARY OF THE EXPERIMENTAL KINEMATIC VARIABLES . S ETTINGS : P1 GAIT SELECTION PARAMETER , P2 ACTIVATION PARAMETER , P3 CPG
LPG S
COUPLING STRENGTH PARAMETER , P4 POSTURE PARAMETER , P5 STEERING PARAMETER . M EASUREMENTS : τ GAIT PERIOD , η GAIT PERIODICITY RATIO
( FROM THE LEG STRAINS εi (t)), σ 2 [θ(t)] PITCH VARIANCE , σ 2 [φ(t)] ROLL VARIANCE , hẑ(t)i AVERAGE ELEVATION OVER TERRAIN , hδx(t)i AVERAGE
LONGITUDINAL SPEED ( BODY FRAME ), hδy(t)i AVERAGE TRANSVERSE SPEED ( BODY FRAME ), hδΨ(t)i AVERAGE YAW RATE , ∆t DATA LENGTH . “∅”
DENOTES ABSENCE OF SUSTAINED MOTION , “∀” AN IRRELEVANT SETTING . E ACH ROW REPRESENTS A DISTINCT EXPERIMENTAL RUN n; RUNS DENOTED
WITH “ ∗ ” ARE REPETITIONS , WHICH WERE PERFORMED TO EVALUATE REPEATABILITY. C ORRESPONDING DATA FILES AVAILABLE ONLINE [36].

σ 2 [θ(t)] σ 2 [φ(t)] hẑ(t)i hδx(t)i hδy(t)i hδΨ(t)i

n P1 P2 P3 P4 P5 τ /s η ∆t/s
/deg2 /deg2 /mm /(mm s−1 ) /(mm s−1 ) /(deg s−1 )
Ant-like posture, walking straight
1 0.2 1 1 0 0 2.94 0.96 6.87 3.52 57.1 28.9 4.6 0.93 105.9
2 0.3 1 1 0 0 3.12 0.89 5.50 3.34 57.8 26.7 3.7 1.01 100.2
3 0.4 1 1 0 0 3.38 0.91 7.14 3.34 64.7 22.5 2.7 0.89 105.1
4 0.5 1 1 0 0 3.38 0.89 12.60 7.75 69.6 33.0 1.6 -0.14 102.9
5 0.6 1 1 0 0 ∅ ∅ ∅ ∅ ∅ ∅ ∅ ∅ n/a
6 0.7 1 1 0 0 2.31 0.78 9.21 2.81 93.6 34.2 -0.1 1.26 89.2
7 0.8 1 1 0 0 2.31 0.91 7.32 2.88 93.9 49.6 -2.5 -0.14 92.9
8∗ 0.8 1 1 0 0 2.31 0.82 7.01 2.62 93.9 57.7 -8.0 -0.28 92.9
9∗ 0.8 1 1 0 0 2.31 0.88 7.27 2.25 95.1 48.6 -4.4 -0.86 94.7
Cockroach-like posture, walking straight
10 0.2 1 1 1 0 2.94 0.92 1.22 15.43 72.8 24.0 -6.2 0.13 110.8
11∗ 0.2 1 1 1 0 2.94 0.94 0.51 13.40 67.2 23.4 -2.3 0.02 111.0
12∗ 0.2 1 1 1 0 2.94 0.93 0.65 15.81 68.7 19.2 0.5 -0.42 76.2
13 0.3 1 1 1 0 3.19 0.94 0.60 13.57 67.6 24.4 -0.9 0.01 111.7
14 0.4 1 1 1 0 3.31 0.90 0.78 13.64 65.6 20.4 -2.8 0.06 115.4
15 0.5 1 1 1 0 3.38 0.67 0.65 16.32 61.4 18.0 0.3 -0.19 109.7
16 0.6 1 1 1 0 ∅ ∅ ∅ ∅ ∅ ∅ ∅ ∅ n/a
17 0.7 1 1 1 0 2.31 0.75 1.13 13.60 70.1 15.4 -4.5 0.01 101.1
18 0.8 1 1 1 0 2.31 0.67 1.23 15.95 75.2 31.1 -3.7 -0.14 91.4
Ant-like posture, steering
19 0.8 1 1 0 -1 2.31 0.89 2.45 2.64 95.0 28.0 18.2 7.19 94.7
20∗ 0.8 1 1 0 -1 2.31 0.86 3.67 2.49 89.1 27.7 13.6 6.41 95.1
Uncoordinated movement
21 ∀ 1 0 0 0 4.44 0.55 13.17 4.12 66.6 3.1 -1.1 -0.01 120.8
22∗ ∀ 1 0 0 0 6.72 0.52 20.26 6.60 61.6 1.9 -1.3 0.11 115.1

cockroach-like posture was associated to larger roll oscillation stabilization of the initial transient all combinations of settings
regardless of the gait (e.g., Fig. 8g). We additionally observed provided a straight trajectory, with the exception of the runs
that, compared to the situation for the slower wave and tetra- with non-zero steering parameter P5 , which successfully gen-
pod gaits, walking in the faster tripod gait and ant-like posture erated a continuous rotation of the body along the yaw axis.
lead to the robot spontaneously adopting a more elevated These data confirm that all gaits and both postures provided
position, plausibly also owing to a more even load distribution effective locomotion. This was in spite of the fact that control
between the legs; the gradual emergence of this feature is was purely open-loop, namely there was no feedback to
well-evident in Fig. 7h. As expected, under the condition adjust posture, heading etc. The trajectories in world frame
of uncoordinated movement, disordered attitude changes and coordinates x̂ and ŷ corresponding to the cases considered
particularly large pitch swings arose (Fig. 9g), accompanied above are shown in Fig. 7d, Fig. 8d and Fig. 9d.
by the largest observed instability of body elevation, manifest It should nevertheless be noted that the repeated runs per-
with the robot cyclically raising itself and falling back towards formed for four representative settings indicated that the
the ground as visible in Fig. 9h. repeatability of the kinematic measurements was incomplete:
Consideration of the average longitudinal and transverse (i.e., as reported in Table III, the relative error was generally on
forward and sideways) speeds in the body frame, referred to the order of 10%, but at times reached substantially higher
as δx(t) and δy(t), confirmed some features of the gaits in levels, particularly for transverse speed and rotation, especially
line with pre-established biological observations, namely that when the corresponding values were small. A systematic
in the ant-like posture, the tripod gait delivered considerably reproducibility evaluation was out of scope for the present
faster locomotion than the wave and intermediate gaits; in the study, but given the high stability of the pattern generation
cockroach-like posture, the differences between gaits were less (η ≈ 1 as indicated above, and negligible variability of
marked and the speed was on average lower. The transverse period τ ), it appears plausible that the error was dominated
speed, ideally zero, was always considerably lower than the by mechanical factors, particularly slippage of the feet against
longitudinal speed, i.e. |δy(t)| ≈ |δx(t)|/10. As expected, the floor due to the limited contact area.
for uncoordinated movement very limited effective locomotion
was observed. Consideration of the yaw rate δΨ(t) further
indicated that, besides relatively contained fluctuations, after

2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2018.2799145, IEEE Access

Fig. 6. Representative gaits obtained as a function of the gait selection parameter P1 , while setting P2 = P3 = 1 and P4 = P5 = 0, and corresponding
frame sequences acquired with the suspended robot (i.e., legs not contacting terrain). a) Wave gait; b,d,e,f) Intermediate gaits; c) Tetrapod (metachronal) gait;
g) Tripod gait. In the gait diagrams, representing the CPG output signals, black and white areas respectively denote the swing and stance phases, and the
frame times are indicated by red bars. The red halo in each frame represents the robot posture in the previous frame shown. CPG: Central Pattern Generator.
Corresponding video files available online [36].

IV. D ISCUSSION cell in a cellular non-linear network, while fixing the tibia-
femur joint to a predetermined angle [4], [21], [34]. Some
A. Hierarchical architecture controllers are “hybrid” in that they include a bio-inspired
Hexapod robots realistically modeling insects generally pos- CPG implemented in the form of a circuit, but the translation
sess three degrees of freedom per leg, raising the challenge of its outputs into leg actuator signals is achieved by means
of translating each CPG output into multiple signals, suitable of conventional robotics techniques, namely inverse kinematic
for driving the actuators coupled to the coxa-body, femur- modeling [43], [44]. Other controllers simply implement linear
coxa and tibia-femur joints [3]–[5], [8], [12]. In particular, mapping between the CPG outputs and the leg joints [45].
physically realizing a given “global” gait pattern (representing To the authors’ knowledge, only a minority of studies have
the phase relationships between the legs) requires identifying realized truly hierarchical controllers, wherein each leg is
a set of suitable “local” patterns (i.e., joint trajectories for associated to a local network of nodes which receive the
the individual legs); this is a computational problem of con- corresponding CPG input and process it in a non-linear manner
siderable complexity, which insects appear to solve seamlessly depending on additional parameters [22], [46], [47].
when realizing the canonical gaits in a highly adaptive manner In this work, a fully hierarchical approach was adopted,
depending, for example, on surface roughness and inclination wherein each leg is associated to a ring-like structure com-
[1], [2]. prising three nodes that effectively constitute a distinct pattern
In the majority of existing controllers, a non-hierarchical generator, driven by the corresponding CPG node to which
approach is used and the individual leg actuators are directly it is unidirectionally coupled. This level of complexity was
driven based on multiple outputs generated by each oscillator instrumental to attaining the demonstrated level of flexibility,
instanced in the CPG; for example, the coxa-body and femur- for example by making it possible to seamlessly shift between
coxa joints can be mapped to the x and y outputs of each two completely different locomotion schemes (or postures),

2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2018.2799145, IEEE Access

Fig. 7. Experimental data acquired for walking according to the tripod gait, in the ant-like posture (run n = 7; P1 = 0.8, P2 = P3 = 1 and P4 = P5 = 0).
a) Gait diagram, representing the CPG output signals; black and white areas respectively denote the swing and stance phases. b) Example output signals from
the CPG and LPG nodes of leg R1; the joints α, β and γ are specified in Fig. 5b. c) Corresponding leg trajectory in joint angles space [α, β, γ], arbitrary units.
d) Robot trajectory in world frame coordinates x̂ and ŷ; the triangle denotes starting location, the circles denote locations at 10 s intervals. e) Representative
frame sequence, spanning approximately one gait period. f) Femur strain signals εi (t). g) Pitch and roll signals θ(t) and φ(t). h) Body elevation over terrain
signal ẑ(t). CPG: Central Pattern Generator. LPG: Local Pattern Generator. Corresponding video and data files available online [36].

one wherein the power stroke is exclusively delivered by the how tightly the LPG dynamics are driven by the CPG, i.e.
coxa-body joints and another wherein the same is primarily “extrinsic”, as opposed to internally-generated independently
delivered by the tibia-femur joints. of the CPG, i.e. “intrinsic”. Low values of this parameter
It is also noteworthy that, in comparison to some of the other lead to desynchronization, visible in the form of uncoordi-
hierarchical controllers cited above, the present network was nated movement resembling, at least at the surface, thrashing
appreciably smaller and therefore more suitable for physical as observed in insects during some forms of neuroreceptor
implementation as an analog circuit. As previously mentioned, blockade (e.g., insecticide poisoning), or during attempts to
even though VLSI- and FPGA-based CPG implementations become untrapped [1], [2]. A similar capability is also present
have been proposed, to the authors’ knowledge this is also the in other circuits based on chaos control, but the difference
first FPAA-based implementation of a complete CPG; in the is that in such cases the transition to chaos is driven by
area of bio-inspired robotics, custom FPAAs have previously absence or alteration of environmental feedback to the CPG,
been used only for robot trajectory planning and to implement whereas in the present case it is internally managed by a
a leaky-integrate-and-discharge type oscillator [27], [48], [49]. specific parameter [22]. The ability to selectively decouple
Another notable feature of the proposed controller is the the activity of individual legs from the CPG (not explicitly
presence of a coupling strength parameter explicitly setting demonstrated in study, but clearly implied in the architecture

2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2018.2799145, IEEE Access

Fig. 8. Experimental data acquired for walking according to the wave gait, in the cockroach-like posture (run n = 10; P1 = 0.2, P2 = P3 = P4 = 1 and
P5 = 0). a) Gait diagram, representing the CPG output signals; black and white areas respectively denote the swing and stance phases. b) Example output
signals from the CPG and LPG nodes of leg R1; the joints α, β and γ are specified in Fig. 5b. c) Corresponding leg trajectory in joint angles space [α, β, γ],
arbitrary units. d) Robot trajectory in world frame coordinates x̂ and ŷ; the triangle denotes starting location, the circles denote locations at 10 s intervals.
e) Representative frame sequence, spanning approximately one gait period. f) Femur strain signals εi (t). g) Pitch and roll signals θ(t) and φ(t). h) Body
elevation over terrain signal ẑ(t). CPG: Central Pattern Generator. LPG: Local Pattern Generator. Corresponding video and data files available online [36].

of the controller) can provide a richer framework to study provisioned to allow enabling and inhibiting the CPG, and
failure compensation compared to the situation wherein a leg even obtaining approximate reverse patterns. It was demon-
is entirely disabled (e.g., effectively amputated), as was done strated that the CPG responds satisfactorily to both gradual and
for chaos-based control in Ref. [47]. step changes of the gait selection parameter, effectively setting
walking speed, as well as to step changes of the activation
B. Generalized gait parameter. In a practical application such parameters would
The notion of a generalized gait introduced in Ref. [34] was be controlled dynamically, i.e. as P1 (t) . . . P5 (t), by a system
implemented with recourse to three sigmoidal membership playing the role of higher-level areas to purposefully drive the
functions peaking at low, mid and high values of a single gait robot. Another noteworthy aspect of the proposed controller
selection parameter, following principles similar to those of is that the discrete-time implementation based on switched-
neuro-fuzzy membership. As discussed above, depending on capacitor circuits ensures that all parameters are effectively
this parameter the CPG produces a variety of gaits including orthogonal to the clock rate; as a consequence, the frequency
not only the “canonical” wave, tetrapod (or metachronal) and at which a given gait is generated can be controlled completely
tripod gaits but also a number of intermediate ones [1], [2], independently of all other parameters, effectively decoupling
[4], [34]. Besides gait selection, the activation parameter was gait and locomotion speed.

2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2018.2799145, IEEE Access

Fig. 9. Experimental data for uncoordinated movement (run n = 22; P1 = 0.2, P2 = 1, P3 = P4 = P5 = 0). a) Gait diagram, representing the CPG output
signals, which are irrelevant in this case since the LPGs were uncoupled. b) Example output signals from the CPG and LPG nodes for leg R1; the joints α, β
and γ are specified in Fig. 5b. c) Corresponding leg trajectory in joint angles space [α, β, γ], arbitrary units. d) Robot trajectory in world frame coordinates
x̂ and ŷ; the triangle denotes starting location, the circles denote locations at 10 s intervals. e) Representative frame sequence, spanning approximately 7 s. f)
Femur strain signals εi (t). g) Pitch and roll signals θ(t) and φ(t). h) Body elevation over terrain signal ẑ(t). CPG: Central Pattern Generator. LPG: Local
Pattern Generator. Corresponding video and data files available online [36].

As expected and summarized in Table III, there were several had more similar stability between the two postures, with
differences between the gaits and postures. The tripod gait the exception of greater roll in the cockroach-like posture.
(high P1 ), commonly observed in insects in association with Uncoordinated movement resulted in the most irregular femu-
fast locomotion for example during escape, accordingly pro- ral strain signals, very limited locomotion and highest pitch
vided maximum speed [1], [2]; its period was spontaneously variance, even though rolling was smaller compared to the
lower than that of the wave gait and intermediate gaits, cockroach-like posture. These results highlight the viability of
representing an emergent property since no time constant or the controller in a wide range of configurations. It should be
filter frequency parameters were altered, only the connectivity noted, however, that the observed kinematic differences reflect
pattern was changed. This gait provided the highest speed not only emergent features of the controller network itself, but
and best stability, according to both pitch/roll variance and are also the result of the characteristics of the robot mechanics
femural strains, when reproduced in the ant-like posture, (e.g. weight distribution, maximum joint torque etc.) and of
which also spontaneously resulted in a higher body elevation. the values assigned to the large number of constant parameters
Contrariwise, the wave gait (low P1 ), commonly observed in specified in Tables A.I and A.III.
insects in association to situations requiring slow locomotion
such as foraging [1], [2], provided the lowest speed and

2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2018.2799145, IEEE Access

C. Role of the non-linearity knowledge are unique in literature, in that its hierarchical ar-
chitecture delivers a notable level of posture and coordination
Recent work has shown that chaotic oscillators harbor
versatility, because the operation of the CPG and that of the
substantial potential for bio-inspired robotics, in that they
LPGs are both parameterized.
allow generating a broad range of gait patterns depending on a
Third, in terms of adaptiveness with respect to walking over
single control parameter, under the assumption that feedback
slopes and irregular terrains, existing controllers predicated
from the environment is available [22], [43], [47]. In the
on connectionist and chaos-control approaches are remark-
initial study where our oscillator was introduced, an extended
ably successful [17], [19], [22], [44], [51]. Assessing the
region of the control parameter space was explored aiming
performance of the present architecture in this regard was
to obtain periodic and chaotic collective oscillation of a ring
beyond the scope of the work. Furthermore, standardized
network, and in particular to elicit the phenomenon of remote
testing conditions and measures for comparing this aspect of
synchronization, which occurs close to the transition between
performance are yet to be defined.
the two phases [26]. Here, however, the non-linear oscillator
Fourth, external controllability, intended as the ability to
operated in a very different region and frequency band, and
purposefully alter activity via changing parameter settings,
neither transition to chaos nor bifurcations were observed. In
has also been extensively demonstrated for connectionist and
other words, the non-linear nature of the oscillator was simply
chaos-control approaches, and the characteristics of the present
exploited to obtain different stable limit cycles and phase
controller appear in line with existing literature [4], [22], [34].
relationships, unlike for example Ref. [22] where it played
However, given the diversity of approaches there is again no
a deeper generative role.
consensus regarding the testing conditions, and future work
Rather than being fully emergent from the dynamics, the
should explicitly compare the available controllers in terms of
transitions between the canonical gaits were for the most part
latency and stability in response to control parameter changes.
directly driven by the connectivity pattern under the influence
Fifth, in terms of stability of the intrinsically-generated gait
of the gait selection parameter; in this sense, the present con-
patterns, the available literature offers largely anecdotal ev-
troller is more founded on synchronization patterns than on the
idence in the form of illustrations rather than quantitative
dynamical properties of the individual oscillators. However,
measurements (e.g. Ref. [4]). In the present study a measure of
additional emergent gaits were also observed. Furthermore,
periodicity, the periodicity ratio η, was introduced and allowed
emergent signatures of non-linear dynamics were nevertheless
explicitly confirming the high stability of controller output.
observed around the point of transition between the wave-
Sixth, in terms of ability to drive a physical robot in a
like gaits and the tripod gait, i.e. for P1 ≈ 0.6: at that point,
kinematically-stable manner, to the authors’ knowledge no
the transition between the two gaits was clearly hysteretic,
comprehensive evaluations have as yet been performed. A
hence a first-order one. In the same region, disordered patterns
prerogative of the present study is that the issue was addressed
could be generated, breaking the periodicity observed for the
explicitly via telemetry of body rotations and structural strains.
regular gait patterns, and metastability was also evident. These
The results indicate a moderate level of stability where,
results are noteworthy as they closely recall observations of
as previously discussed, certain combinations of gaits and
gait transitions in insects and other animal species [38], [50].
postures are highly viable whereas others were reproduced
sub-optimally. Standardized testing conditions and measures
D. Remarks on performance
are again lacking and an explicit comparison of the existing
Unbiased performance comparison between CPG-based approaches would represent an important future contribution.
controllers is presently complicated by the diversity of ap- In this regard, the present approach, which involved calcu-
proaches and absence of standardized criteria, nevertheless it lating the periodicity ratio η from the leg strains εi (t) and
is appropriate to further juxtapose the proposed controller with considering it in conjunction with the pitch and roll variances
the existing literature in this regard. σ 2 [θ(t)] and σ 2 [φ(t)], may have general value.
First, compared to the initial works wherein the notion of Lastly, the walking speed of the robot was on the order of
generalized gait was introduced, the proposed controller offers 0.02 m/s to 0.06 m/s. While this value is strongly influenced
similar or superior capability as regards the range of canonical by mechanical aspects and parameter settings and thus not
and intermediate gaits which can be generated, with the truly representative of controller performance, we note that
added advantage that gait selection is controlled by a single it is overall comparable with other hexapod robots driven by
parameter [4], [34]. Compared to chaos-based controllers, the bio-inspired controllers. In particular, it closely overlaps ex-
diversity of gaits also appears overall comparable, however it perimental measurements of controllers based on connectionist
should be noted that the present controller did not generate approaches (e.g., 0.03 m/s to 0.05 m/s in Ref. [4]); however,
separate slow and fast wave gaits [22]. It is also noteworthy compared to some experiments with chaos-based controllers
that the gait summary parameter ζ introduced in the present the maximum speed was lower and the difference between
study has general usefulness for evaluating and comparing the fastest and slowest gaits was accordingly more constrained
gait-generating architectures. (e.g., 0.03 m/s to 0.14 m/s in Ref. [22]).
Second, in terms of implementation of the generated gaits, by
design the present controller has features which to the authors’

2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2018.2799145, IEEE Access

E. Directions for future research node is an FPAA-based oscillator was introduced. Besides its
Since the proposed controller is inherently well-suited for inherent flexibility, the controller features a hierarchical archi-
rapid and ongoing reconfiguration, future work will leverage tecture which confers a high level of versatility in gait, posture
this capability. For example, several previous studies have and coordination. It was demonstrated that the resulting large
demonstrated the possibility of delivering stable locomotion, number of low-level parameters can be handled through con-
heading and attitude even in presence of instabilities and tol- structing the corresponding expressions as the superposition of
erances in the actuators and mechanics, irregular surfaces and a minimal number of canonical gaits and postures, as specified
slopes, by means of suitable feedback signals, and attempts via five high-level parameters. The controller was deployed to
should be made to replicate such results using the proposed a physical robot, and in order to confirm its viability, a range
controller [17], [19], [44], [51]. In fact, maneuverability over of experimental measures were considered, which also warrant
irregular terrain is one of the primary strengths of hexapod future consideration as means of comparing controllers in this
robots compared to rovers, and the main limitation of this area.
initial study is that the controller was only demonstrated in A PPENDIX
an open-loop configuration, which inherently prohibits loco-
motion in presence of irregularities or obstacles. Addressing As for determination of the single oscillator parameters,
this important issue will likely require the introduction of given the absence of a formal approach we proceeded
additional parameters beyond the five ones considered in this iteratively for the B parameters, specified in Table A.I and
study, to be adjusted based on inertial data and terrain contact further described in Table A.II, in the following manner. First,
feedback, allowing controlling pitch and roll independently of we considered only the case of extrinsic activity (P3 = 1) and
all other aspects; since Tables A.I and A.III contain all infor- thus fixed B1 = B2 = B6 = 0: 1) given the ant-like posture
mation required to map the CPG outputs to the corresponding (P4 = 0), in which all legs operate identically, we fixed
legs, such additional parameters will enter the expressions B4 = B5 = B8 = 0 and adjusted B3 , B7 and B9 in order
contained in these tables. Furthermore, it should be noted to obtain robust synchronization and the phase relationships
that the constant values set in this initial study were chosen between the α, β and γ outputs necessary to implement viable
empirically and as such may require optimization for improved leg trajectories (examples of which can be found in Ref.
performance; this could be attained, for example, using genetic [4]), 2) given the cockroach-like posture (P4 = 1), we then
algorithms as previously done in Ref. [18]. adjusted B4 , B5 and B8 to obtain the desired differentiation
To realize more complex and realistic behaviors the high-level for the front and hind legs, 3) because parameters B3 , B7 and
control parameters P1 . . . P5 , rather than being statically set B9 affect both postures, they had to be iteratively adjusted
for each experimental run, should be driven dynamically as trying to find settings acceptable for both. In steps 1) and 2)
P1 (t) . . . P5 (t) by afferences from “higher” circuits capable, each free parameter was initially set to either 1 or −1, then
for example, of exhibiting learning based on the delivery of adjusted. Second, we considered the case of intrinsic activity
reward such as in Ref. [52], and of performing purposeful (P3 = 0) and adjusted B1 , B2 and B6 until sustained leg
tasks based on visual and other multisensory input such as in swings were elicited, with features similar to the circle-like
Refs. [43], [46], [53]. movements observed biologically. As the C parameters were
A controller such as the present one, wherein significant not yet defined, these steps were completed with the controller
versatility is condensed into a handful of parameters, also has network deployed on the physical hardware, but evaluating
inherent potential in the field of brain-computer and brain- signal plots rather than materially driving the robot.
machine interfaces, wherein usability is knowingly inversely As regards the C parameters, specified in Table A.III
related to the number of commands and variables that need and further described in Table A.IV, we proceeded
to be decoded [54]. Importantly, the present work allows con- in a similar sequence. First, we considered only the
trol in a seamless, continuous manner, without necessitating case of extrinsic activity (P3 = 1) and thus fixed
the decoding of discrete commands, and as such may have C2 = C4 = C6 = C8 = C11 = C12 = C15 = C16 =
specific advantages in terms of operability and learning curve C18 = C20 = C22 = C24 = C26 = C29 = 0: 1)
compared to more traditional interfaces [55]. Future research given the ant-like posture (P4 = 0), we further fixed
will therefore include real-time control with ongoing decoding C1 = C7 = C10 = C19 = C25 = C28 = 0 and adjusted C3 ,
of all or a subset of the five parameters from electroencephalo- C5 , C9 , C13 , C14 , C17 , C21 , C23 and C27 in order to obtain
graphic signals or other bio-signals, implemented for instance adequate swing on all leg axes, 2) given the cockroach-like
by means of directly mapping power density at specific sites posture (P4 = 1), we then adjusted C1 , C7 , C10 , C19 , C25
or cortical regions to the parameters, or with recourse to sparse and C28 to obtain adequate power stoke delivery with the
activity representations [56]. front and hind legs, 3) because C13 , C14 and C17 affect both
postures, they had to be iteratively adjusted trying to find
V. C ONCLUDING REMARKS settings acceptable for both. Second, we considered the case
This work aimed to address the issues of flexibility and of intrinsic activity (P3 = 0) and adjusted C2 , C4 , C6 , C8 ,
availability associated with implementing CPGs by means of C11 , C12 , C15 , C16 , C18 , C20 , C22 , C24 , C26 and C29 until
custom integrated circuits. A novel CPG network wherein each the corresponding movements were performed adequately.

2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2018.2799145, IEEE Access

TABLE A.I
H ARDWIRED LPG NODE INPUT CONNECTIVITY AND CORRESPONDING WEIGHTS ( I . E ., INPUT GAINS ) FOR EACH LPG NODE . T HE GAINS G1 , G2 AND G3
ARE DETERMINED BY PARAMETERS P3 AND P4 , WHICH ARE HIGH - LEVEL SETTINGS AS SPECIFIED IN TABLE I; B1 . . . B9 ARE CONSTANTS , SEE TEXT
FOR DESCRIPTION . VOLTAGES vL AND vC DENOTE THE TWO EXTERNAL INPUTS OF EACH NODE , ONE CONNECTED TO ANOTHER NODE IN THE SAME
LPG AND THE OTHER TO THE CPG NODE WHICH DRIVES THE LPG, AND WEIGHTED RESPECTIVELY BY THE GAINS G1 AND G2 ; G3 DETERMINES THE
LOOP GAIN WITHIN EACH NODE ( SEE F IGS . 3 B AND C ). F OR SUCCINCTNESS , P30 = 1 − P3 AND P40 = 1 − P4 . CPG: C ENTRAL PATTERN G ENERATOR .
LPG: L OCAL PATTERN G ENERATOR .

For node vL vC G1 G2 G3
L1/α −4v2,L1/γ v5,L1/C B1 P30 P3 B2 P30 + B3 P3 P40
R2/α −4v2,R2/γ v5,R2/C B1 P30 P3 B2 P30 + B3 P3
L3/α −4v2,L3/γ v5,L3/C B1 P30 P3 B2 P30 + B3 P3 P40
R1/α −4v2,R1/γ v5,R1/C B1 P30 P3 B2 P30 + B3 P3 P40
L2/α −4v2,L2/γ v5,L2/C B1 P30 P3 B2 P30 + B3 P3
R3/α −4v2,R3/γ v5,R3/C B1 P30 P3 B2 P30 + B3 P3 P40
L1/β −v5,L1/α v5,L1/C −P30 P3 (1 + B4 P4 ) B5 P3 P4
R2/β −v5,R2/α v5,R2/C −P30 P3 0
L3/β −v5,L3/α v5,L3/C −P30 P3 (1 + B4 P4 ) B5 P3 P4
R1/β −v5,R1/α v5,R1/C −P30 P3 (1 + B4 P4 ) B5 P3 P4
L2/β −v5,L2/α v5,L2/C −P30 P3 0
R3/β −v5,R3/α v5,R3/C −P30 P3 (1 + B4 P4 ) B5 P3 P4
L1/γ −4v2,L1/β v5,L1/C B6 P30 + B7 P3 P40 −B8 P3 P4 B9 P3 P40
R2/γ −4v2,R2/β v5,R2/C B6 P30 + B7 P3 0 B 9 P3
L3/γ −4v2,L3/β v5,L3/C B6 P30 + B7 P3 P40 B8 P3 P4 B9 P3 P40
R1/γ −4v2,R1/β v5,R1/C B6 P30 + B7 P3 P40 −B8 P3 P4 B9 P3 P40
L2/γ −4v2,L2/β v5,L2/C B6 P30 + B7 P3 0 B 9 P3
R3/γ −4v2,R3/β v5,R3/C B6 P30 + B7 P3 P40 B8 P3 P4 B9 P3 P40

TABLE A.II
D ESCRIPTION OF THE B PARAMETERS , WHICH INFLUENCE LPG NODE INPUT CONNECTIVITY AS SPECIFIED IN TABLE A.I. LPG: L OCAL PATTERN
G ENERATOR .

Parameter Control target Applicable configuration(s) Target legs Target joints

B1 Within-LPG γ
α coupling Uncoordinated movement All α
B2 Internal loop gain Uncoordinated movement All α
B3 Internal loop gain Walking in ant-like (front, hind legs) or any posture (middle legs) All α
B4 CPG
LPG coupling Walking in cockroach-like posture Front and hind β
B5 Internal loop gain Walking in cockroach-like posture Front and hind β
B6 Within-LPG β
γ coupling Uncoordinated movement All γ
B7 Within-LPG β
γ coupling Walking in ant-like (front, hind legs) or any posture (middle legs) All γ
B8 CPG
LPG coupling Walking in cockroach-like posture Front and hind γ
B9 Internal loop gain Walking in ant-like (front, hind legs) or any posture (middle legs) All γ

TABLE A.III
O UTPUT GAINS FOR LPG TO SERVO - MOTOR CONNECTIVITY ( SEE F IGS . 3 B AND C ). T HE GAINS G4 AND G5 ARE DETERMINED BY PARAMETERS P3 , P4
AND P5 , WHICH ARE HIGH - LEVEL SETTINGS AS SPECIFIED IN TABLE I; C1 . . . C29 ARE CONSTANTS , SEE TEXT FOR DESCRIPTION . F OR SUCCINCTNESS ,
P30 = 1 − P3 , P40 = 1 − P4 , P50 = 1 − R(−P5 ) AND P500 = 1 − R(P5 ), WHERE R(x) IS THE RAMP FUNCTION . LPG: L OCAL PATTERN G ENERATOR .

For node G4 G5
L1/α C1 P3 P4 C2 P30 + C3 P3 P40 P50
R2/α 0 C12 P30 + C13 P3 P500
L3/α C19 P3 P4 C20 P30 + C21 P3 P40 P50
R1/α C1 P3 P4 C2 P30 + C3 P3 P40 P500
L2/α 0 C12 P30 + C13 P3 P50
R3/α C19 P3 P4 C20 P30 + C21 P3 P40 P500
L1/β C4 P3 + C5 P3 P40
0 C6 P30 + C7 P3 P4 P50
R2/β C14 C15 P30
L3/β C22 P30 + C23 P3 P40 C24 P30 + C25 P3 P4 P50
R1/β C4 P30 + C5 P3 P40 C6 P30 + C7 P3 P4 P500
L2/β C14 C15 P30
R3/β C22 P30 + C23 P3 P40 C24 P30 + C25 P3 P4 P500
L1/γ C8 P30 + C9 P3 P40 + C10 P3 P4 C11 P30
R2/γ C16 P30 + C17 P3 C18 P30
L3/γ C26 P30 + C27 P3 P40 + C28 P3 P4 C29 P30
R1/γ C8 P30 + C9 P3 P40 + C10 P3 P4 C11 P30
L2/γ C16 P30 + C17 P3 C18 P30
R3/γ C26 P30 + C27 P3 P40 + C28 P3 P4 C29 P30

2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2018.2799145, IEEE Access

TABLE A.IV
D ESCRIPTION OF THE C PARAMETERS , WHICH INFLUENCE LPG TO SERVO - MOTOR CONNECTIVITY AS SPECIFIED IN TABLE A.III. LPG: L OCAL
PATTERN G ENERATOR .

Parameter Control target Applicable configuration(s) Target legs Target joints

C1 Position offset Walking in cockroach-like posture Front α
C2 Coupling coefficient for LPG direct output Uncoordinated movement Front α
C3 Coupling coefficient for LPG direct output Walking in ant-like posture Front α
C4 Coupling coefficient for LPG mixed output Uncoordinated movement Front β
C5 Coupling coefficient for LPG mixed output Walking in ant-like posture Front β
C6 Coupling coefficient for LPG direct output Uncoordinated movement Front β
C7 Coupling coefficient for LPG direct output Walking in cockroach-like posture Front β
C8 Coupling coefficient for LPG mixed output Uncoordinated movement Front γ
C9 Coupling coefficient for LPG mixed output Walking in ant-like posture Front γ
C10 Coupling coefficient for LPG mixed output Walking in cockroach-like posture Front γ
C11 Coupling coefficient for LPG direct output Uncoordinated movement Front γ
C12 Coupling coefficient for LPG direct output Uncoordinated movement Middle α
C13 Coupling coefficient for LPG direct output Walking in any posture Middle α
C14 Coupling coefficient for LPG mixed output All Middle β
C15 Coupling coefficient for LPG direct output Uncoordinated movement Middle β
C16 Coupling coefficient for LPG mixed output Uncoordinated movement Middle γ
C17 Coupling coefficient for LPG mixed output Walking in any posture Middle γ
C18 Coupling coefficient for LPG direct output Uncoordinated movement Middle γ
C19 Position offset Walking in cockroach-like posture Rear α
C20 Coupling coefficient for LPG direct output Uncoordinated movement Rear α
C21 Coupling coefficient for LPG direct output Walking in ant-like posture Rear α
C22 Coupling coefficient for LPG mixed output Uncoordinated movement Rear β
C23 Coupling coefficient for LPG mixed output Walking in ant-like posture Rear β
C24 Coupling coefficient for LPG direct output Uncoordinated movement Rear β
C25 Coupling coefficient for LPG direct output Walking in cockroach-like posture Rear β
C26 Coupling coefficient for LPG mixed output Uncoordinated movement Rear γ
C27 Coupling coefficient for LPG mixed output Walking in ant-like posture Rear γ
C28 Coupling coefficient for LPG mixed output Walking in cockroach-like posture Rear γ
C29 Coupling coefficient for LPG direct output Uncoordinated movement Rear γ

These steps were completed first observing the movement gratefully acknowledges funding by the World Research Hub
of individual legs with the robot suspended and making no Initiative, Institute of Innovative Research, Tokyo Institute of
terrain contact, then observing actual walking. Technology, Tokyo, Japan.
Due to the large number of parameters and their interactions,
the adjustment process was extremely labor intensive and
required several weeks of experimental work. The above R EFERENCES
indications represent only general guidelines, and in practice [1] D. M. Wilson, “Insect walking,” Annu. Rev. Entomol., vol. 11, pp. 103–
a large number of adjustments were performed involving 122, 1966.
simultaneous tuning of both B and C parameters, aiming [2] F. Delcomyn, “Insect walking and robotics,” Annu. Rev. Entomol., vol
49, no. 1, pp. 51–70, Jan. 2004.
to maximize stability as well as kinematic realism; in [3] B. Webb, “What does robotics offer animal behaviour?,” Anim. Behav.,
particular, further tuning was needed to avoid “tetanus” vol. 60, no. 5, pp. 545–558, Nov. 2000.
due to competition between multiple configurations given [4] M. Frasca, P. Arena, and L. Fortuna, Bio-Inspired Emergent Control of
Locomotion Systems.. Singapore: World Scientific, 2004.
intermediate settings of P3 and P4 . It is due to this difficulty [5] M. R. Clark, G. T. Anderson, and R. D. Skinner, “Coupled Oscillator
that other studies, even when gaits are generated by a CPG, Control of Autonomous Mobile Robots,” Auton. Robots., vol. 9, no. 2,
almost universally resort to simple mappings or conventional pp. 189–198, Sep. 2000.
[6] M. Schilling, T. Hoinville, J. Schmitz, and H. Cruse, “Walknet, a bio-
kinematic models; while considerably less demanding to inspired controller for hexapod walking,” Biol. Cybern., vol. 107, no. 4,
implement, such approaches are however also less biologically pp. 397–419, Aug. 2013.
plausible [43]–[45]. [7] H. Cruse, T. Kindermann, M. Schumm, J. Dean, and J. Schmitz,
“Walknet - a biologically inspired network to control six-legged walk-
ing,” Neural Netw., vol. 17, n. 7-8, pp. 1435–1447, Oct. 1998.
[8] C. Ferrell, “A comparison of three insect-inspired locomotion con-
ACKNOWLEDGEMENT trollers,” Rob. Auton. Syst., vol. 16, no. 2, pp. 135–159, Dec. 1995.
[9] U. Bässler, “On the definition of central pattern generator and its sensory
This work was supported in part by JSPS under Grant control,” Biol. Cybern., vol 54, no. 1, pp. 65–69, May 1986.
KAKENHI 26112004 and Grant KAKENHI 17H05903 [10] F. Delcomyn, “Neural Basis of Rhythmic Behavior in Animals,” Science,
vol. 210, no. 4469, Oct. 1980.
and in part by the Center of Innovation Program from the [11] E. Marder, and D. Bucher, “Central pattern generators and the control
Japan Science and Technology Agency, JST. Realization of rhythmic movements,” Curr. Biol., vol. 11, no. 23, pp. 986–996, Nov.
of the hexapod robot hardware and related electronics was 2001.
[12] A. J. Ijspeert, “Central pattern generators for locomotion control in
personally funded by L. Minati. In addition to employment by animals and robots: A review,” Neural Netw., vol. 21, no. 4, pp. 642–653,
the Polish Academy of Sciences, Kraków, Poland, L. Minati May 2008.

2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2018.2799145, IEEE Access

[13] J. Yu, M. Tan, J. Chen, and J. Zhang, “A Survey on CPG-Inspired [35] Understanding Anadigm® Configurable Analog Modules (CAMs),
Control Models and System Implementation,” IEEE Trans. Neural Netw. Anadigm, Inc. Mesa AZ, technical documentation available at http:
Learn. Syst., vol. 25, no. 3, pp. 441–456, Mar. 2014. //www.anadigm.com/ apps/PR021100-0024.pdf.
[14] A. Billard, A. J. Ijspeert, “Biologically inspired neural controllers [36] Publicly Available Data, accessed on Sep. 11, 2017. [Online]. Available:
for motor control in a quadruped robot”, in Neural Networks, 2000. http://www.lminati.it/listing/2017/b/.
IJCNN 2000, Proceedings of the IEEE-INNS-ENNS International Joint [37] B. Boashash, “Estimating and interpreting the instantaneous frequency
Conference on, vol. 6, pp. 637-641. of a signal. II. Algorithms and applications,” Proc. IEEE, vol. 80, no.
[15] M. A. Lewis, R. Etienne-Cummings, M. J. Hartmann, Z. R. Xu, and 4, pp. 540–568, Apr. 1992.
A. H. Cohen, “An in silico central pattern generator: silicon oscillator, [38] S. Fujiki, S. Aoi, T. Funato, N. Tomita, K. Senda, and K. Tsuchiya,
coupling, entrainment, and physical computation,” Biol. Cybern., vol. “Hysteresis in the metachronal-tripod gait transition of insects: a mod-
88, no. 2, pp. 137–151, Feb. 2003. eling study,” Phys Rev E, vol. 88, no. 1, pp. 012717, Jul. 2013.
[16] K. Matsuoka, “Mechanisms of frequency and pattern control in the [39] J. A. Acebrón, L. L. Bonilla, C. J. Pérez Vicente, F. Ritort, and R.
neural rhythm generators,” Biol. Cybern., vol. 56, no. 5, pp. 345–353, Spigler, “The Kuramoto model: A simple paradigm for synchronization
Jul. 1987. phenomena,” Rev. Mod. Phys., vol. 77, no. 1, pp. 137–185, Apr. 2005.
[17] R. Campos, V. Matos, and C. Santos, “Hexapod locomotion: A non- [40] H. Hyyti, and A. Visala, “A DCM Based Attitude Estimation Algorithm
linear dynamical systems approach,” IECON 2010 - 36th Annual Con- for Low-Cost MEMS IMUs,” Intl. J. Navigat. Observ., vol. 25, pp.
ference on IEEE Industrial Electronics Society, Glendale AZ, 2010, pp. 503814, Nov. 2015.
1546–1551. [41] H. Bay, A. Ess, T. Tuytelaars, and L. Van Gool, “SURF: Speeded Up
Robust Features,” Comput. Vis. Image Underst., vol. 110, no. 3, pp.
[18] W. Li, W. Chen, X. Wu and J. Wang, “Parameter tuning of CPGs for
346–359, Jun. 2008.
hexapod gaits based on Genetic Algorithm,” 2015 IEEE 10th Conference
[42] P. H. S. Torr, and A. Zisserman, “MLESAC: A New Robust Estimator
on Industrial Electronics and Applications (ICIEA), Auckland, 2015, pp.
with Application to Estimating Image Geometry,” Comput. Vis. Image
45–50.
Underst., vol. 78, no. 1, pp. 138–156, Apr. 2000.
[19] P. Arena, L. Fortuna, and M. Frasca, “Attitude control in walking [43] P. Milička, P. Petr, and J. Faigl, “On Chaotic Oscillator-Based Central
hexapod robots: an analogic spatio-temporal approach,” Int. J Circ. Pattern Generator for Motion Control of Hexapod Walking Robot.”, ITAT
Theor. App., vol. 30, n. 2-3, pp. 349–362, Mar. 2002. 2016 Proceedings, CEUR Workshop Proceedings, Tatranske Matliare,
[20] P. Arena, and L. Fortuna, “Analog cellular locomotion control of Slovakia, vol. 1649, pp. 131–137.
hexapod robots,” IEEE Control Systems, vol. 22, no. 6, pp. 21–36, Dec. [44] W. H. Chen, G. J. Ren, J. h. Wang, and D. Liu, “An adaptive locomotion
2002. controller for a hexapod robot: CPG, kinematics and force feedback,”
[21] P. Arena, L. Fortuna, M. Frasca, and L. Patane, “A CNN-based chip for Sci. China Inf. Sci., vol. 57, no. 11, pp. 1–18, Nov. 2014.
robot locomotion control,” IEEE Trans. Circuits Syst. I, Reg. Papers, [45] H. Yu, W. Guo, J. Deng, M. Li, and H. Cai, “A CPG-based locomotion
vol. 52, no. 9, pp. 1862–1871, Sep. 2005. control architecture for hexapod robot,” 2013 IEEE/RSJ International
[22] S. Steingrube, M. Timme, F. Worgotter, and P. Manoonpong, “Self- Conference on Intelligent Robots and Systems, Tokyo, 2013, pp. 5615-
organized adaptation of a simple neural circuit enables complex robot 5621.
behaviour,” Nat. Phys., vol. 6, no. 3, pp. 224–230, Jan. 2010. [46] L. A. Fuente, M. A. Lones, A. P. Turner, L. S. Caves, S. Stepney, and
[23] A. Buscarino, L. Fortuna, M. Frasca, G. Muscato, “Chaos does help A. M. Tyrrell, “Adaptive robotic gait control using coupled artificial
motion control”, International Journal of Bifurcation and Chaos, 17(10), signalling networks, hopf oscillators and inverse kinematics,” 2013 IEEE
3577-3581, 2007. Congress on Evolutionary Computation, Cancun, 2013, pp. 1435–1442.
[24] A. Pikovsky, M. Rosenblum, and J. Kurths, Synchronization: A Uni- [47] G. Ren, W. Chen, S. Dasgupta, C. Kolodziejski, F. Wörgötter, and P.
versal Concept in Nonlinear Sciences. Cambridge, UK: Cambridge Manoonpong, “Multiple chaotic central pattern generators with learning
University Press, 2003. for legged locomotion and malfunction compensation,” Inf. Sci., vol.
[25] L. Minati, “Experimental synchronization of chaos in a large ring of 294, pp. 666–682, Feb. 2015.
mutually coupled single-transistor oscillators: phase, amplitude, and [48] S. Koziol, P. Hasler, and M. Stilman, “Robot path planning using Field
clustering effects,” Chaos, vol. 24, no. 4, pp. 043108, Dec. 2014. Programmable Analog Arrays,” 2012 IEEE International Conference on
[26] L. Minati, “Remote synchronization of amplitudes across an experimen- Robotics and Automation, Saint Paul MN, 2012, pp. 1747–1752.
tal ring of non-linear oscillators,” Chaos, vol. 25, no. 12, pp. 123107, [49] S. Koziol, R. Wunderlich, J. Hasler, and M. Stilman, “Single-Objective
Path Planning for Autonomous Robots Using Reconfigurable Analog
Dec. 2015.
VLSI,” IEEE Trans. Syst., Man., Cybern., Syst., vol. 47, no. 7, pp. 1301–
[27] F. Li, A. Basu, C. H. Chang, and A. H. Cohen, “Dynamical Systems
1314, July 2017.
Guided Design and Analysis of Silicon Oscillators for Central Pattern
[50] J. J. Collins, and I. Stewart, “Hexapodal gaits and coupled non-linear
Generators,” IEEE Trans. Circuits Syst. I, Reg. Papers, vol. 59, no. 12,
oscillator models,” Biol. Cybern., vol. 68, no. 4, pp. 287–298, Feb. 1993.
pp. 3046–3059, Dec. 2012.
[51] Y. Ambe, T. Nachstedt, P. Manoonpong, F. Wörgötter, S. Aoi and F. Mat-
[28] Q. Chen, J. Wang, S. Yang, Y. Qin, B. Deng, and X. Wei, “A real- suno, “Stability analysis of a hexapod robot driven by distributed non-
time FPGA implementation of a biologically inspired central pattern linear oscillators with a phase modulation mechanism,” 2013 IEEE/RSJ
generator network,” Neurocomputing, vol. 244, pp. 63–80, Jun. 2017. International Conference on Intelligent Robots and Systems, Tokyo,
[29] T. S. Hall, C. M. Twigg, J. D. Gray, P. Hasler, and D. V. Anderson, Japan, 2013, pp. 5087-5092.
“Large-Scale Field-Programmable Analog Arrays for Analog Signal [52] E. Arena, P. Arena, R. Strauss, and L. Patané, “Motor-Skill Learning
Processing,” IEEE Trans. Circuits Syst. I, vol. 52, no. 11, pp. 2298– in an Insect Inspired Neuro-Computational Control System,” Front.
2307, Nov. 2005. Neurorobot., vol. 11, pp. 12, Mar. 2017.
[30] E. Pierzchala, G. Gulak, L. Chua, and A. Rodriguez-Velazquez, Field- [53] J. H. Barron-Zambrano, C. Torres-Huitzil, and B. Girau, “Perception-
Programmable Analog Arrays. New York, NY: Springer, 2013. driven adaptive CPG-based locomotion for hexapod robots,” Neurocom-
[31] A. Basu, S. Brink, C. Schlottmann, S. Ramakrishnan, C. Petre, S. Koziol, puting, vol. 170, pp. 63–78, Dec. 2015.
F. Baskaya, C. M. Twigg, and P. E. Hasler, “A floating-gate based field [54] M. Akcakaya, B. Peters, M. Moghadamfalahi, A.R. Mooney, U. Orhan,
programmable analog array,” IEEE J. Solid-State Circuits, vol. 45, no. B. Oken, D. Erdogmus, and M. Fried-Oken, “Noninvasive brain-
9, pp. 1781–1794, Sep. 2010. computer interfaces for augmentative and alternative communication,”
[32] E. K. F. Lee, and P. G. Gulak, “A CMOS field-programmable analog IEEE Rev. Biomed. Eng., vol. 7, pp. 31–49, 2014.
array,” IEEE J. Solid-State Circuits, vol. 26, no. 12, pp. 1860–1867, Dec. [55] L. Minati, N. Yoshimura and Y. Koike, “Hybrid Control of a Vision-
1991. Guided Robot Arm by EOG, EMG, EEG Biosignals and Head Move-
[33] A. Buscarino, L. Fortuna, M. Frasca, and G. Sciuto, A Concise Guide ment Acquired via a Consumer-Grade Wearable Device,” in IEEE
to Chaotic Electronic Circuits. New York, NY: Springer, 2014. Access, vol. 4, pp. 9528–9541, 2016.
[34] P. Arena, L. Fortuna, M. Frasca, and G. Sicurella, “An adaptive, self- [56] Y. Nakanishi, T. Yanagisawa, D. Shin, H. Kambara, N. Yoshimura, M.
organizing dynamical system for hierarchical control of bio-inspired Tanaka, R. Fukuma, H. Kishima, M. Hirata, and Y. Koike, “Mapping
locomotion,” IEEE Trans. Syst. Man. Cybern. B Cybern., vol. 34, no. 4, ECoG channel contributions to trajectory and muscle activity prediction
pp. 1823–1837, Aug. 2004. in human sensorimotor cortex,” Sci Rep., vol. 7, pp. 45486, Mar. 2017.

2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.