Good Posture, Good Balance · idea is generalized to attain balance and posture control for a...

1070-9932/16©2016IEEE• IEEE ROBOTICS & AUTOMATION MAGAZINE • MARCh 201622

This article provides a theoretical and thorough experimental comparison of two distinct posture control approaches: 1) a fully model-based control approach and 2) a biologically inspired approach

derived from human observations. While the ro-botic approach can easily be applied to balancing in three-dimensional (3-D) and multicontact (MC) situations, the biologically inspired balancer cur-rently only works in two-dimensional situations but shows interesting robustness properties under time delays in the feedback loop. This is an important feature when considering the signal transmission and processing properties in the human sensorimo-tor system. Both controllers were evaluated in a se-ries of experiments with a torque-controlled humanoid robot (TORO). This article concludes with some suggestions for the improvement of model-based balancing approaches in robotics.

A Neurorobotics ApproachNeurorobotics is a relatively new interdisciplinary research field in which results from neuroscience and robotics are combined. It is inspired by the idea that the way the human brain works is inter-linked with the embodiment, i.e., with the proper-ties of the human body, by which the physical interaction with the environment is performed. In the neurorobotics approach, concepts from neuroscience are validated by robotic

Digital Object Identifier 10.1109/MRA.2015.2507098Date of publication: 8 March 2016

Good Posture, Good Balance

By Christian Ott, Bernd Henze, Georg Hettich, Tim Niklas Seyde, Máximo A. Roa, Vittorio Lippi, and Thomas Mergner

Comparison of Bioinspired and Model-Based Approaches for Posture Control of Humanoid Robots

se

es

aw

ima

ge

©is

toc

kp

ho

to.c

om

/sc

iba

k, r

ob

ot

ima

ge

©is

toc

kp

ho

to.c

om

/kir

st

ypa

rg

et

er

march 2016 • IEEE rOBOTIcS & aUTOmaTION maGaZINE • 23

test beds. Moreover, robotics research can profit from neuro-logical findings on biological solutions. This article is con-cerned with a bipedal postural balance against unforeseen external disturbances, which can be considered as one of the most fundamental skills of legged humanoid robots. Balanc-ing has been well studied in the robotics community as a control problem, either based on simplified template models or based on complex multibody models. Early works on bal-ancing utilized the concept of zero-moment point (ZMP) [1]. The ZMP is a ground reference point that describes the point of action of the ground reaction force (GRF). The ZMP has been applied to the problems of dynamic motion genera-tion and reactive balancing against external disturbances [2]–[5]. Intuitively speaking, a feedback loop on the ZMP implements a compliant behavior for a position-controlled robot. While the definition of the ZMP is general, it has been particularly popular in the context of simplified template models.

Balancing based on joint torque interfaces has been studied in the context of floating base multibody dynamics. Recently, several TOROs have appeared [6]–[9]. Early re-sults on passivity-based balancing were presented in [10], and the relation to human postural balancing strategies was discussed in [11]. There, a desired balancing force acting at the robot’s center of mass (COM) was given by a propor-tional-derivative (PD)-like feedback law. This force was dis-tributed to the contact forces at the robot’s feet and further implemented by mapping the contact forces to the robot’s joint torques. In [12], an extension to simultaneous balance and posture control was presented. Moreover, the force dis-tribution problem was formulated as a general optimization problem, allowing for integration of inequality constraints on the contact forces. These works were motivated by im-plementing a compliant reaction in response to an external disturbance, similar to Cartesian compliance control of robot manipulators.

An exact implementation of the GRF can be achieved by inverse dynamics control under the assumptions that a pre-cise model of the robot dynamics is available, the contacts are perfectly rigid, and additional external disturbances on the robot are known. In [13], the contact force distribution problem was solved by using inverse dynamics control. When inverse dynamics approaches are used, the com-manded torque from a whole-body balancing strategy is also sometimes supported by an additional low-level feed-back loop on the joint position. However, this is not strictly needed, as shown in [14].

In neuroscience, the bipedal postural balance has been studied from a different perspective. The aim is to describe the neural mechanisms underlying human balancing in functional models. It has been found that the balancing is highly adapted to the properties of the human sensorimotor system. When describing human balancing from a neuroro-botics perspective, three aspects stand out.

● First, humans use feedback loops with different time de-lays. The musculoskeletal properties generate a joint

torque almost without time delay (passive loop). Short latency (SL) primitive reflexes via the spinal cord and brainstem occur due to signal transmission and force generation with a time delay of about 80 ms (SL loop). Long latency (LL), context-dependent reactions via the basal ganglia and cerebral cortex require up to 300 ms (LL loop).

● Second, information from several sensory systems, includ-ing the vestibular system, joint proprioception, and vision, is integrated into human balancing. Humans dynamically modify the contribution of each sensory system to ac-count for changes in environmental scenarios and sensor availability.

● Third, humans show a feedback loop gain in their control of balance that is slightly above the minimum to prevent falling [15]. The low loop gain provides the system with high robustness with respect to long neural time delays and with mechanical compliance.A model of human balancing that captures the above-de-

scribed aspects has been proposed by the disturbance estima-tion and compensation (DEC) concept [16]. The DEC concept successfully reproduces human balancing in simula-tions and in experiments using a humanoid robot [17].

The goal of the study presented in this article is to compare a bioinspired posture control mecha-nism with an engineer-ing-inspired mechanism during different external disturbances using a TORO. The comparison analyzes the performance of the controllers in a qualitative and quantitative way with respect to different time delays and sampling rates in the feedback loops. For the comparison, we chose an extended version of the bioinspired controller from [16] and the model-based balancing ap-proach from [12]. The comparison also aims at deriving sug-gestions for improving current model-based balancing controllers for humanoid robots. Figure 1 shows an overview of the experiments presented in this article. Besides balanc-ing in the presence of disturbances on firm and stationary ground, we consider a tilting surface, a horizontal accelera-tion of the feet, and a compliant floor. The comparison fo-cuses on balancing in the sagittal (lateral) plane, which can be handled by both controllers. The model-based balancing controller also allows a compliant behavior in the frontal plane and an extension to multiple contacts, including the hands. The evaluation of the model-based balancing control-ler also covers these generalized cases, which are currently not supported by the bioinspired controller.

Model-Based Bipedal Balancing In this article, we consider a model-based balancing algo-rithm that extends the controller from [12]. The basic idea of

Both controllers were

evaluated in a series of

experiments with a torque-

controlled humanoid robot.

• IEEE ROBOTICS & AUTOMATION MAGAZINE • MARCh 201624

the balancing approach is to control the robot’s total COM x R3! in such a way that it shows a compliant response with respect to external disturbances. It is well known that for every mechanical multibody system, the COM dynamics fulfills the equation

,x g f fm m extc= + +p (1)

where m is the total mass, g is the gravity vector, fc is the sum of all external forces at the robot’s end effectors in con-tact, and fext is an external (disturbance) force. If fc is considered as a control input and the virtual equilibrium po-sition of the COM is denoted by ,xd such a compliant behav-ior is easily achieved via the control law g Kf mc t= + ( )x x D xd t- - o for some positive definite stiffness and damping matrices Kt and .Dt The question is then how to realize the total force fc via the robot’s actuators. For that, we have to take the force distribution to the end effectors in con-tact into account. In this section, this conceptually simple

idea is generalized to attain balance and posture control for a complete humanoid.

Given a desired equilibrium x Rd3! for the COM posi-

tion and a desired orientation ( )R SO 3d ! for the trunk, a desired control wrench Fd acting at the COM is computed via the compliance control law

( )( , ) .x x

RFg

Dx K

Rm0d

t d

k d~ x= - -

-oc c cm m m (2)

Here, R and ~ denote the orientation and the angular ve-locity of the trunk, respectively, both measured by the onboard inertial measurement unit (IMU). Damping is implemented by the matrix D. Kt and kx denote a translational stiffness and the torque from a virtual rotational spring between R and .Rd The torque is computed according to ( K2k r #x ed=- +

) ,Kr e with the scalar part d and the vector part e of the qua-ternion describing the error in the rotation ( )R Rd

T [18]. A for-mation similar to (2) can also be used for controlling the angular momentum instead of the trunk orientation [19].

Figure 1. The setup of the experiments for the comparison of the model-based and the bioinspired controllers. (a) Transient behavior. (b) Redundancy of the MC controller. (c) 3-D balancing. (d) Tilt disturbance in the support surface. (e) Translational acceleration disturbance in the support surface. (f) Compliant support surface.

(a) (b) (c)

(d) (e) (f)


This desired wrench Fd is then distributed to the support-ing contacts at the feet. Therefore, we consider the mapping between the stacked vector ( )W Fi= of contact wrenches Fi and the total wrench F via the grasp matrix G, i.e., .F GW= The desired contact wrenches Fi are then computed based on a constrained minimization of the error

|| | | | | | |

. . ,minW W W F GW

AWs t a

defd

#

c= - + -) ^ h

(3)

where the desired wrench is implemented by means of a soft constraint using a large weighting parameter .1&c Further-more, additional constraints AW a# can be imposed on the optimization problem to consider unilaterality, friction, and location of the center of pressure for each contact.

To realize the desired contact wrenches ,Fi we map these contact wrenches to the corresponding joint torques .dx For an exact implementation of the contact forces, the whole mul-tibody robot dynamics should be considered. As a simpler so-lution, we can also consider a kinetostatic mapping based on the relevant Jacobian matrices, which can be derived by con-sidering the embedding of the COM dynamics (1) within the complete multibody dynamics [12], [19]. This leads to

,J F J Fd rT

r lT

lx = + (4)

where Jr and J l denote the Jacobian matrices for the right and left foot with respect to the COM location .x In this way, the compli-ance (2) is implemented for the whole body dy-namics without requiring an explicit control or measurement of the con-tact forces. More details on the derivation of the controller are presented in [12].

The resulting control scheme does not require any measurement of con-tact forces (neither at the feet nor at other locations on the robot), and can be made robust against un-certainties in the ground geometry by utilizing a state measurement based on the IMU data. It can readily be extended to a larger set of contact points involving the arms and legs to realize more com-plex MC interaction tasks, as shown in Figure 2.

To illustrate the features of this controller, two experi-ments were conducted with the TORO [8]. In the first exper-iment, the robot balances on its legs while the arms are not in contact. In this configuration, a number of arbitrary distur-bances are applied to the robot, demonstrating the capability of this controller to handle 3-D balancing [Figure 1(c)]. The full experiment can be seen on the video accompanying this publication in IEEE Xplore. The second ex-periment illustrates the ability to exploit multiple contacts for balancing, including the resulting re-dundancy, by solving the optimization problem mentioned previously. In this scenario, the robot is supporting itself with all four end effectors [see Figure 1(b)]. In the default pose, the controller distributes the weight of the robot to all four end effectors by generating vertical contact forces of ap-proximately ffoot, r = N372 and Nf 340foot, l = for the feet, while the arms carry a weight of Nf 35hand, r = and

Nf 39hand, l = (Figure 3). As soon as the robot is manually pushed down ( s),t 3 7= - it tries to counteract the distur-bance by increasing the contact forces. When the robot is pulled up ( s),t 8 14= - the optimization in (3) prevents

Figure 2. The notation and block diagram for the model-based balancing controller with an example of a multicontact situation.

Fd

Rd

Fr

Fl

mg

xd

X

YZ

ForceDistribution

Force Mapping

Torque Control

RobotDynamics

Object ForceGeneration

q

Torque-Controlled Robot [8]

IMUKinematics

xd, Rd

FixdFd xm

xx, R

The bipedal postural

balance has been

studied from a different

perspective.


negative contact forces at the hands by a lower limit thresh-old that avoids losing the contact.

Design of the DEC Controller The DEC controller describes human balancing of biped stance during external disturbances, being used here in a simplified form. Originally, it focused on sagittal plane balanc-ing around the ankle joints, which allows simplifying the human biomechanics to a single-inverted pendulum (SIP) [16], [20]. The basic structure of the DEC concept is shown in Figure 4(a). The three loops are 1) passive loop, 2) SL joint angle proprioceptive feedback loop, and 3) LL disturbance compensation loop. Combined, the passive and SL loops form a servo mechanism. With appropriately adjusted parameters and in the absence of external disturbances, it ensures that the actual joint position matches the desired joint position given by a set point signal. The SL and LL loops use a neural PD controller [box neural controller (NC)] to provide the motor command for producing active joint torque. In this article, the neural time delays of SL and LL are represented in one-lumped time delay (box ) .es tD The LL loop commands the servo to compensate external disturbances estimated on the basis of sensory signals. Four types of external disturbances can be distinguished: 1) field forces, such as gravity, 2) contact forces, such as a push against the body, 3) support surface tilt, and 4) support surface translational acceleration. The DEC controller allows superimposing a disturbance on other distur-bances and on voluntary movements.

Human data on ankle joint balancing have successfully been reproduced in computer simulations using the DEC concept and through its implementation in a special purpose SIP robot, called Posturob I [16], [17]. The concept was fur-ther developed as a modular control system, where the com-plexity of the control structure scales linearly with the number of degrees of freedom (DoF) [21]. The control struc-ture originally described for the ankle joint control is applied to each of the controlled joints in the form of DEC control modules, where each module controls 1 DoF. The reconstruction of the external disturbances affecting a joint is performed through an exchange of sensory signals between adjacent modules. The information coming from the vestib-ular system is down channeled through the modules and fused with the corresponding proprioceptive signals. Thus, the control reconstructs, for example, the orientation of the foot to compensate support surface tilts, providing robust-ness with respect to uneven and compliant ground. The sen-sory interaction between modules creates an internal model of the robot’s kinematics. Depending on the behavioral task, the module can be set to control the COM position of all the links above the controlled link, the joint angle, or the orienta-tion in space of the supported link. The modular concept was tested in a DEC controller that included ankle and hip

Figure 3. The exploitation of the redundancy in a multicontact scenario where the contact forces are normalized with the weight of the robot.

-0.02-0.01

00.010.02

00.0250.05

0.075

0.1 Right Left Lower Limit

Right Left

0 2 4 6 8 10 12 14 160.4

0.450.5

0.550.6

Dx z

(m

)f H

and,

i/(

migm)

f Foo

t, i/

(migm)

Time t (s)

(c)

0 2 4 6 8 10 12 14 16Time t (s)

(b)

0 2 4 6 8 10 12 14 16Time t (s)

(a)

RobotPulled Up

RobotPushed Down

Figure 4. The DEC controller with (a) a schematic model of the DEC concept and (b) a modular control architecture used in the experiments.

DisturbanceEstimators

Set Pointes.DtNC

Passive PD

VestForceProp

MultibodyDynamics

SL

LL

Hip Module

Knee Module

Ankle Module

SensorySignals

Knee Torque

AnkleTorque

Hip Torque

SensorySignals

Vest

Hip Prop

Ankle Prop

Knee Prop

(b)

(a)


joints successfully reproducing human balancing responses in model simulations and in a special purpose 2-DoF robot called Posturob II [22].

In this article, the DEC controller was transferred to the control of TORO, which meant entering unknown ground. The controller was extended to include the knee joints by adding one more module to the ankle and hip modules [Fig-ure 4(b)]. In humans, knee joint contributions tend to be minor with sagittal plane disturbances [23], [24]. Qualitative-ly, the knees tend to be locked in an extended position dur-ing moderate disturbances with the system in steady state, while strong transient disturbances tend to evoke a transient knee bending. Not knowing yet exactly the human knee re-sponses during transient disturbances, the knee control module was set to the task of joint angle stabilization with a bent position of 39° from the straight position. Missing DEC model settings for transient responses, we resorted to the DEC model established for steady-state balancing. The ankle module was set to maintain a whole body COM lean of 0.9° forward, and the hip module was set to maintain an upper body orientation of 0.9° forward.

The implementation of the estimators of contact forces and the support surface translation into TORO did not bring significant improvements, and it increased the danger of con-trol instability in some cases. Therefore, only the tilt and grav-ity estimators were active in the current study, having previously observed that these estimators make the major contribution to balance control [22].

In each control module, the gravitational torque is estimat-ed and compensated in the general form

( ),sinw m ghgrav, grav , ,i i i i BS ix a=t (5)

where the index i addresses a joint (i.e., hip joint is one, knee joint is two, and ankle joint is three), ,BS ia represents the space orientation of the COM of all segments above the ith joint, wgrav, i is a weighting factor, and mi and hi represent the total mass and COM height of all segments above the ith joint, re-spectively. For small angles, we assume that hi is constant and

( ) .sin , ,BS i BS ia a= To obtain ,,BS ia the tilt disturbance, i.e., the space orientation of the supporting link ( ),space, ja is required. This is estimated from vestibular ( )vesta and proprioceptive ( )prop, ia signals in the form

.space, vest prop,j ii

j

1

1

a a a= -=

-

t / (6)

The combination of vestibular and proprioceptive signals allows controlling the system in space coordinates. Previous studies suggested the presence of thresholds and weighting factors in the estimators and a servo loop gain to reproduce human data. Since the aim of this article was not to reproduce human results but to compare the two balance controllers, the servo loop gain ws was set to 0.8. The thresholds were not ap-plied, and the weighting factors were set close to unity ( . )w 0 95c = to achieve a disturbance compensation close to the ideal one. The total control torque ix applied at each joint (ankle, knee, and hip) at time t consists of an instantaneous

(passive) joint-level PD action with setpoint q ,d i and a delay affected servo loop, including a PD action on :,BS ia

( ) ( ) ( )( )

( ( ) ( )) .

t t t tK q q D qw

w K D

, ,

, , , ,

, est, grav, , ,

, , , , , , ,

i p i a i

p i p i i d i p i i

a i c i i PD i p i

PD i s a i d i BS i a i d i BS i

x x x

x

x x x x x

x a a a a

D= + -

=- - -

=- - + -

= - + -

ot t

o o

(7)

The control torque ,,a ix representing the SL and LL feed-back loops, is affected by a time delay .tD The effects of the inner passive control ac-tion ,p ix as well as other estimated disturbances

est, ixt are taken into ac-count in .,a ix The values of the passive controller gains, K ,p i and ,D ,p i were set to 15% of the active controller gains, K ,a i and

,D ,a i according to values observed in humans [15]. For technical reasons, the

vesta signal from the IMU and the input into the de-rivative part of the con-trollers were processed with a first-order low pass-filter with the cutoff frequencies of 30 and 5 Hz, respectively.

Hypotheses and Concept of Experiments

Experimental Hypotheses The two control approaches presented in the previous sections aim at balancing against external perturbations. They have been derived from different perspectives. The model based balancer (MC controller) uses a complete dynamical model of the humanoid robot and aims at controlling the force distribu-tion while achieving a compliant behavior. On the other hand, the DEC controller was derived from observations of the human balancing behavior. It employs a modular control concept for the ankle, knee, and hip joints. Moreover, it applies a fast joint-level feedback loop representing the human muscle properties in addition to the DEC. This fast joint-level control action plays an important role in human balancing due to the different time delays of the neural feedback loops.

This motivates the hypothesis that the passive loop of the DEC controller leads to increased robustness against time delay and slow controller sampling. However, it is expected that the modular control approach of the DEC controller leads to stronger dynamic couplings in the transient response than in the case of the model-based controller that uses a cen-tralized control structure.

Implementation We conducted several experiments for both controllers using the TORO [8], [25]. In the current version, TORO has 27 DoF (excluding the hands), a height of 1.74 m, and a weight of 76 kg.

The combination

of vestibular and

proprioceptive signals

allows controlling

the system in space

coordinates.


The estimation of the state of the COM was obtained using in-formation coming from internal sensors, namely the onboard IMU and the kinematic information of the robot. Since the DEC controller is designed specifically for balancing in the sagittal plane, the MC controller was artificially restricted to move only

in the same plane to pro-vide a fair comparison be-tween both concepts. For the DEC concept, the con-trol action ,p ix of the pas-sive loop would ideally be implemented through a local low-level joint con-troller working at a high sampling rate, separate from the additional high-level full-state feedback controller. Our implemen-

tation of this low-level passive loop uses the same real-time system as the high-level controller. While low- and high-level commands were generated by the high-level controller environ-ment, the former maintained a 0-ms artificial delay and 1,000-Hz sampling rate throughout all trials.

All experiments have been performed with the default set of parameters listed in Tables 1 and 2, unless otherwise stated. These parameters were tuned independently for the nominal case (no time delay and fast controller sampling) on flat ground.

Experimental Comparison

Transient Behavior The first experiment is focused on the transient behavior generated by the two controllers to illustrate the conceptual

differences between both approaches. The robot was manu-ally pushed down and to the front before being released, and the trajectory of the COM in the sagittal plane was recorded. This process was repeated 14 times for both the DEC-and the MC controllers. To highlight the controller’s intrinsic be-havior, we reduced (for this experiment only) the stiffness and the damping in the knees to 10 and 70% of their original values (provided in Table 2), respectively, thus allowing a larger deviation of the robot from the equilibrium configuration (especially along the vertical axis).

The recorded trajectories are displayed in Figure 5, where a typical trajectory is also highlighted for a better under-standing. The MC controller moves the COM along an al-most linear trajectory connecting the perturbed position with the equilibrium position and displays a moderate over-shoot. In fact, as the MC controller is based on the system’s COM state feedback, it generates a virtual spring and damper between the current and the desired COM location, thus generating a trajectory almost along a straight line. In con-trast, the DEC approach can be considered as a controller based on a modular joint feedback. It generates an indepen-dent torque command for each joint module without inter-modular communication. The trajectories in Figure 5 show a mutual obstruction between modules, displaying a circular overshoot. Here, the knee modules command an extension of the knee pitch joints. As the upper body is also perturbed to the front, an extension of the knees is counterproductive for the ankle pitch joints, which try to move the COM back into the equilibrium position. This effect can reduce the overall performance of the DEC controller in certain configurations, as the one examined in this experiment.

Tilt Disturbance in the Support Surface One of the characteristics of the DEC algorithm is its robust-ness against controller delays, as these are inherent to ner-vous system control. Introducing artificial delays in the

Figure 5. The COM trajectories for the MC and the DEC controller after a perturbation from equilibrium in the sagittal plane. Two consecutive dots correspond to positions obtained with a time difference of 100 ms. The color of the highlighted trajectories fades from red into blue over time. (a) The DEC controller’s discrete modules follow conflicting objectives, producing trajectories with circular overshoot. (b) The MC controller’s whole state feedback-based algorithm produces almost linear trajectories.

-0.03 0 0.03 0.06

−0.03

0

0.03

-0.03 0 0.03 0.06

−0.03

0

0.03

Start Start

Goal Goal

DEC MC

Dx (m)

Dz

(m)

Dz

(m)

Dx (m)

(a) (b)

The DEC concept successfully

reproduces human

balancing in simulations

and in experiments using

a humanoid robot [17].

Table 1. The default parameterization of the MC controller.

Parameter Value

Kt diag( , N/m, , N/m, , N/m)1 500 1 500 3 000

D diag( . Ns/m, . Ns/m, Ns/m,0 25 0 15 0 f

Nms/rad, Nms/rad, Nms/rad)15 17 10

Kr diag( Nm, Nm, Nm)200 100 200

Table 2. The default parameterization of the DEC controller.

Parameter Ankle Knee Hip

K ,a i m gh3 3 m gh2 2 m gh1 1

D ,a i . m gh0 30 3 3 . m gh0 20 2 2 . m gh0 20 1 1

K ,p i . m gh0 15 3 3 . m gh0 15 2 2 . m gh0 15 1 1

D ,p i . m gh0 05 3 3 . m gh0 05 2 2 . m gh0 05 1 1


high-level commands of our controllers and asserting the re-spective robustnesses were, therefore, identified as an inter-esting aspect with regard to biomimetics. From a technical perspective, the sampling rate for the execution of the high-level controllers becomes an important factor, as it restricts the complexity of the underlying computations. We, there-fore, chose the high-level sampling rate as the second indi-vidual parameter to be varied.

We compared the behavior of the controllers when the robot is subjected to a tilt disturbance on the support surface [Figure 1(d)]. As the DEC controller features a passive loop, which can be implemented directly within the joints’ low-lev-el controllers, we created two versions of the DEC controller, namely, DEC without passive loop (DEC )PL ffo- and DEC with passive loop (DEC ).PL on- The delay and sampling rates of the low-level commands were hereby not altered and re-mained at 0 ms and 1,000 Hz, respectively.

The system was placed on a wooden rocker board that pivoted at its center around the normal to the sagittal plane [Figure 1(d)]. The total pivoting angle was 4°. The board was kept horizontal by using a support block, and the back of the robot’s heels was aligned with the pivoting axis. The system’s

COM was located in front of the axis. The support block was then removed manually, causing the board to tilt due to the robot’s weight, while the robot swayed forward. For both the MC and DEC controllers, the time delay was increased from 0 ms, until the point where the system became unstable. In a similar way, the sampling rate was de-creased (in the absence of delay) starting from 1,000 Hz, until the point where the system became unsta-ble. The obtained trajec-tories for the COM pitch angle with respect to the normal to the sagittal plane passing through the robot’s ankles are shown in Figure 6(a) and (b).

The MC controller was able to reject much higher delays than the DEC controller when the passive loop was inactive (Figure 6). While the former remained stable until 70 ms of time delay, the latter was unable to withstand 15 ms of delay, as the ankle joints shut off due to active torque limit violation. However, activating the passive loop in the DEC controller

−20246

−20246

0 1 2 3 4 5 6 7 8−2

0246

-20246

0 ms 50 ms 100 ms 120 ms

0 ms 40 ms 70 ms 80 ms

−2024

−20246

0 1 2 3 4 5 6 7 8

−20246

−20246

6 1,000 Hz 100 Hz 20 Hz 10 Hz

1,000 Hz 100 Hz 20 Hz10 Hz 5 Hz

1,000 Hz 100 Hz 20 Hz10 Hz 5 Hz

aS

uppo

rt (

°)D

EC

PL–

Off

aC

OM

(°)

DE

CP

L–O

na

CO

M (

°)M

Ca

CO

M (

°)

aS

uppo

rt (

°)D

EC

PL–

Off

aC

OM

(°)

DE

CP

L–O

na

CO

M (

°)a

CO

M (

°)

Time t (s)

(a)

Time t (s)

(b)

0 ms 10 ms 15 ms

Figure 6. The COM pitch angle trajectories for a tilt disturbance in the support surface with (a) different high-level control delays and (b) different sampling rates. The top row shows the instant where the support of the surface is removed (black) and the measured pitch angle of the feet (gray). The DEC without passive loop controller (DEC )PL ffo- is outperformed by the MC controller ( mst 10maxD = versus ms,t 70maxD = respectively). Activation of a low-level passive loop in the joints (DEC with passive loop, DEC )PL no- increases DEC performance ( ms)t 100maxD = above the performance of the MC controller. The same trend is observed for different sampling rates of high-level commands: the low-level passive loop brings the DEC without passive loop controller ( Hz)f 20sampling min, = almost on par with the MC controller ( Hz).f 10sampling,min =

The passive and SL

loops form a servo

mechanism.


significantly improved the performance. Joint failure was pre-vented and the system remained stable until 100 ms of time delay, increasing robustness against high-level delay even above the MC controller. The passive loop allowed the system to react to the disturbance instantaneously, passive stiffness and damping counteracting plantar flexion. Therefore, the

impact of the impulse dis-turbance was reduced before the high-level con-troller became active to stabilize the robot.

A similar trend was observed when consider-ing the high-level sam-pling rate, as shown in

Figure 6(b). Activating the passive loop in the DEC control-ler increased robustness against low sampling rates from 20 to 10 Hz, making it comparable to the MC controller. The data obtained from these experiments confirms that incor-porating a low-level passive loop with the original high-level controller can have a positive influence on the controller’s overall robustness.

Acceleration Disturbance in the Support Surface To confirm the increase in robustness against high-level de-lays and sampling rate limitations by the passive loop, we introduced a different perturbation. The system was placed on a linear motion platform powered by a linear motor, as shown in Figure 1(e). The platform was displaced 0.04 m within 0.5 s, starting from rest, moving the robot back-wards. The resulting COM pitch angle trajectories for the MC controller and the two variants of the DEC controller are provided in Figure 7(a) and (b), when different time de-lays and sampling rates are considered.

The DEC without the passive loop controller has a higher robustness for this type of perturbation [Figure 7(a)], as it became unstable at 15 ms of time delay for the tilting surface, and now it is stable up to 30 ms of time delay (at 40 ms there was an emergency shutdown due to instability). However, it still performed worse than the MC controller, which was ca-pable of reaching 40 ms of time delay (although it could stand up to 70 ms of time delay with the tilting surface). Again, an activation of the passive loop increased the perfor-mance of the DEC controller, rejecting a time delay of 80 ms, which outperforms the MC controller. Note that the DEC

−20246

−20246

0 1 2 3 4 5 6 7 8−2

0246

−2024

−20246

0 1 2 3 4 5 6 7 8

−20246

6 1,000 Hz 100 Hz 20 Hz 10 Hz0 ms 20 ms 30 ms 40 ms

0 ms 40 ms 80 ms 90 ms

0 ms 20 ms 40 ms 50 ms 1,000 Hz 100 Hz 20 Hz 10 Hz

1,000 Hz 100 Hz 20 Hz10 Hz 5 Hz

-0.05

0

0.05

0.1

x Sup

port (

°)D

EC

PL–

Off

aC

OM

(°)

DE

CP

L–O

na

CO

M (

°)M

Ca

CO

M (

°)

DE

CP

L–O

ffa

CO

M (

°)D

EC

PL–

On

aC

OM

(°)

MC

aC

OM

(°)

Time t (s)

(a)

Time t (s)

(b)

-0.05

0

0.05

0.1

x Sup

port (

°)

Figure 7. The COM pitch angle trajectories for a translation disturbance in the support surface with (a) different high-level control delays and (b) different sampling rates. The top row shows the commanded and measured translation for the support surface (black and gray, respectively). The DEC without passive loop controller (DEC )PL ffo- is outperformed by the MC controller ( mst 03maxD = versus ms,t 40maxD = respectively). Activation of a low-level passive loop in the joints (DEC )PL on- increases DEC with passive loop controller performance ( ms)t 80maxD = above the MC controller performance. The same trend is observed for the different sampling rates of high-level commands: the passive loop brings the DEC with passive loop controller performance ( Hz)f 10sampling, min = beyond the MC controller performance ( Hz).f 20sampling, min =

Humans use feedback

loops with different

time delays.


controller with passive loop rejected up to 100 ms with the tilting surface.

The consideration of different sampling rates, shown in Figure 7(b), supported these findings. The MC and DEC without passive loop controllers remained stable until 20 Hz of sampling rate. With the tilting surface, the DEC without passive loop controller reached the same 20 Hz, and the MC controller was more robust as it could reject 10 Hz. The DEC with passive loop reached down to 10 Hz, the same sampling rate it reached with this experi-ment as before.

Compliant Support Surface This section investigates the ability of the MC and the DEC controllers to reject an impulse disturbance while balancing on a compliant support surface (with no time delay for the control signals). The robot was placed with both feet on one to three layers of gym mats, and after-ward it was hit by a mass of 5 kg, as shown in Figure 1(f). The robot received the impact at the hip, and it was in-tended to create a motion mainly in the sagittal plane. One of the biggest challenges of balancing on a compliant support surface is that the robot is not stable by default,

and needs to be actively stabilized by a controller. Since the DEC approach only works in the sagittal plane, mo-tions of the robot out of this plane are needed to be addi-tionally stabilized. Two long ropes were stretched to the left and to the right, preventing the robot from tipping over to the side, while having a long enough lever arm to avoid obstructions to the controller performance in the sagittal plane. In contrast, the MC controller is capable of stabilizing the robot in all directions of space. Instead of using ropes, in this case, all the joints in the lower body were activated, thus allowing the system to stabilize itself in the frontal as well as in the sagittal plane.

The trajectories of the COM pitch angles and the foot pitch angles (left and right) are provided in Figure 8. It can be observed that for both controllers, the amplitude of COM pitch angle sway is significantly smaller than the am-plitude of foot pitch sway. Thus, the ankle joints work very hard toward reducing the overall COM movement, which is smaller for the MC controller than for the DEC control-ler. For the latter, the difference between motion in the left and the right foot is smaller than for the former, which could suggest better adaptation to the underlying control task. However, this could also be simply due to the fact that

Figure 8. (a)–(d) The impulse responses of the DEC controller with passive loop (a, c) and the MC controller (b, d) while balancing on different compliant support surfaces. The top graph shows the COM pitch angle trajectories, while the bottom graph shows the corresponding foot pitch angle trajectories. The index indicates the number of used mattresses. The amplitude of the foot angles is much higher than the COM angles, highlighting the controllers’ effort to stabilize the system. While the DEC with passive loop controller becomes unstable when the robot receives an impulse standing on two layers of compliant support, the MC controller does so on three layers.

-6

-3

0

3

6

9

12

15

0 1 2 3 4 5 6 7 8 9 10-6

-3

0

3

6

9

12

15

Time t (s)

-6

-3

0

3

6

9

12

15

0 1 2 3 4 5 6 7 8 9 10-6

-3

0

3

6

9

12

15

(c)Time t (s)

(d)

(a) (b)

aF

oot (

°)a

CO

M (

°)

aF

oot (

°)a

CO

M (

°)

aCOM,1

aCOM,2

aCOM,3

aFoot,R,1aFoot,L,1aFoot,R,2aFoot,L,2aFoot,R,3aFoot,L,3


the DEC controller lacks DoF in the frontal plane and, therefore, does not have to react to cross-couplings among motions in the two planes. The general performance of the MC controller is better on the compliant surface, as the DEC controller becomes unstable after the impulse pertur-bation when placed on two mattress layers, whereas the MC controller becomes unstable with three layers.

Discussion of the Experimental Results The two control approaches were evaluated by several ex-periments motivated mainly by the disturbances considered in the DEC approach. In addition, we performed experi-ments on a compliant surface, which represents a more gen-

eral uncertainty of the world model. In all cases, we analyzed the robust-ness of the algorithms against time delays and decrease of the sampling rate. For a technical sys-tem, control approaches that are robust against control latency and low sampling rates can re-duce the requirements on the onboard real-time

computer systems, possibly leading to lower-cost solutions.The experimental results show that without time delay

and at 1,000-Hz sampling rate, both controllers are able to stabilize biped stance in the presence of support surface tilts, translations, and on compliant ground. Increasing the time delay or lowering the sampling rate beyond a certain amount led to control instability. The use of a passive stiff-ness and damping in the DEC controller allowed to bear with longer time delays and lower sampling rates. Overall, the various experiments gave consistent results, thus show-ing their generality.

The obtained robot results are comparable with experi-mental results on human balance control. One example is the increase in neural time delays with increasing age [26], which is partly compensated by an increase in passive and active stiffness and damping [27].

Further work is required to improve comparability be-tween controllers when exchanging them across robotic platforms. One example is that the DEC controller in its original Posturob II platform tolerates time delays of more than 120 ms in the ankle joint control during !8° support surface tilt disturbances [22]. In these experiments, the time delay in the hip joint control was set to 70 ms, because this joint is located closer to the spinal cord. Another exam-ple concerns human responses to transient stimuli as used in the current study. Compensation of such stimuli tends to involve cocontraction of agonistic muscle pairs, which tran-siently stabilizes the joints mechanically and, in addition, stabilizes their control by increasing the passive stiffness

[28]. Cocontraction is not yet implemented in the DEC controller. Upon transient stimuli, such as a push or sup-port surface translation, a transient increase in passive stiff-ness due to cocontractions is expected to occur in humans. Furthermore, the DEC controller is designed for balancing in the sagittal plane; a generalization to the lateral plane is work-in-progress. In general, it should be noted that hu-mans are still outperforming the DEC control and the MC control. For a direct comparison, however, it still may be too early.

It is difficult to compare the current results with reports on simulated balancing responses in the literature using models of human balance control, such as the optimal state estima-tion model of [29], because the corresponding robot experi-ments are currently not available.

Summary and Outlook In this article, we provided a theoretical and thorough ex-perimental comparison of two approaches for bipedal bal-ancing: an engineering-inspired model-based balance controller and a bioinspired controller derived from obser-vations of human behavior. Both controllers use a torque interface with the robot. The model-based controller allows for straightforward generalization to balance using multiple end effectors in contact (e.g., hands and feet), and can be applied to balancing in the sagittal and lateral planes. On the other hand, the bioinspired controller utilizes a modu-lar structure and takes the signal transmission properties observed in human posture control into account. It was ex-perimentally confirmed that the DEC controller can be made more robust against time delays in the high-level commands when using a joint-level feedback loop acting at a high sampling rate and without time delay. The experi-mental results suggest some directions for further research. In particular, using a robotic platform may help to obtain a deeper understanding of passive stiffness modulation in human posture control. Furthermore, it would be interest-ing to aim for low-level feedback actions in combination with the model-based controller. In this way, it might be possible to combine the strengths of the model-based con-troller (versatility) and the bioinspired controller (robust-ness against time delay).

Acknowledgment This work was partly supported by the Helmholtz Associa-tion (grant number VH-NG-808) and by the European Commission (FP7-ICT-600698 H2R and H2020-ICT-645097 COMANOID).

References[1] M. Vukobratovic and Y. Stepanenko, “On the stability of anthropomorphic systems,” Math. Biosci., vol. 15, no. 1–2, pp. 1–37, 1972.[2] S. Kajita, F. Kanehiro, K. Kaneko, K. Fujiwara, K. Harada, K. Yokoi, and H. Hirukawa, “Biped walking pattern generation by using preview control of ze-ro-momentpoint,” in Proc. IEEE Int. Conf. Robotics Automation, 2003, pp. 1620–1626.

The LL loop commands

the servo to compensate

external disturbances

estimated on the basis

of sensory signals.


[3] T. Sugihara, Y. Nakamura, and H. Inoue, “Realtime humanoid motion gen-eration through ZMP manipulation based on inverted pendulum control,” in Proc. IEEE Int. Conf. Robotics Automation, 2002, pp. 1404–1409.[4] Y. Choi, D. Kim, Y. Oh, and B.-J. You, “Posture/walking control for human-oid robot based on kinematic resolution of CoM jacobian with embedded mo-tion,” IEEE Trans. Robot., vol. 23, no. 6, pp. 1285–1293, 2007.[5] K. Hu, C. Ott, and D. Lee, “Online iterative learning control of zeromo-ment point for biped walking stabilization,” in Proc. IEEE Int. Conf. Robotics Automation, 2015, pp. 5127–5133.[6] G. Cheng, S.-H. Hyon, J. Morimoto, A. Ude, G. Colvin, W. Scroggin, and S. C. Jacobsen, “CB: A humanoid research platform for exploring neuroscience,” in Proc. IEEE-RAS Int. Conf. Humanoid Robots, 2006, pp. 182–187.[7] J. Kim, Y. Lee, S. Kwon, K. Seo, H. Kwak, H. Lee, and K. Roh, “Develop-ment of the lower limbs for a humanoid robot,” in Proc. IEEE/RSJ Int. Conf. In-telligent Robots Systems, 2012, pp. 4000–4005.[8] J. Englsberger, A. Werner, C. Ott, B. Henze, M. A. Roa, G. Garofalo, R. Burger, A. Beyer, O. Eiberger, K. Schmid, and A. Albu-Schäffer, “Overview of the torque-controlled humanoid robot TORO,” in Proc. IEEERAS Int. Conf. Humanoid Robots, 2014, pp. 916–923.[9] C. S. Knabe, V. Orekhov, M. A. Hopkins, B. Y. Lattimer, and D. W. Hong, “Two configurations of series elastic actuators for linearly actuated humanoid robots with large range of motion,” in Proc. IEEE-RAS Int. Conf. Humanoid Robots, 2014, p. 1096.[10] S. Hyon, J. Hale, and G. Cheng, “Full-body compliant human-humanoid interaction: Balancing in the presence of unknown external forces,” IEEE Trans. Robot., vol. 23, no. 5, pp. 884–898, 2007.[11] S. Hyon, R. Osu, and Y. Otaka, “Integration of multi-level postural balanc-ing on humanoid robots,” in Proc. ICRA, 2009, pp. 1549–1556.[12] C. Ott, M. A. Roa, and G. Hirzinger, “Posture and balance control for biped robots based on contact force optimization,” in Proc. IEEE-RAS Int. Conf. Humanoid Robots, 2011, pp. 26–33.[13] L. Righetti, J. Buchli, M. Mistry, M. Kalakrishnan, and S. Schaal, “Opti-mal distribution of contact forces with inverse-dynamics control,” Int. J. Robot. Res., vol. 32, no. 3, pp. 280–298, 2013.[14] A. Herzog, N. Rotella, S. Mason, F. Grimminger, S. Schaal, and L. Righet-ti, “Momentum control with hierarchical inverse dynamics on a torque-con-trolled humanoid,” Auton. Robot., pp. 1–19. Aug. 2015, DOI: 10.1007/s10514-015-9476-6.[15] R. J. Peterka, “Sensorimotor integration in human postural control,” J. Neurophysiol., vol. 88, pp. 1097–1118, May 2002.[16] T. Mergner, “A neurological view on reactive human stance control,” Annu. Rev. Contr., vol. 34, no. 2, pp. 177–198, 2010.[17] T. Mergner, G. Schweigart, and L. Fennell, “Vestibular humanoid postural control,” J. Physiol. Paris, vol. 103, no. 3–5, pp. 178–194, 2009.[18] F. Caccavale, C. Natale, B. Siciliano, and L. Villani, “Six-DoF impedance control based on angle/axis representations,” IEEE Trans. Robot. Automat., vol. 15, no. 2, pp. 289–299, 1999.[19] G. Garofalo, B. Henze, J. Engelsberger, and C. Ott, “On the inertially de-coupled structure of the f loating base robot dynamics,” in Proc. 8th Vienna Int. Conf. Mathematical Modelling MATHMOD, 2015, pp. 322–327.[20] C. Maurer, T. Mergner, and R. J. Peterka, “Multisensory control of human upright stance,” Exp. Brain Res., vol. 171, no. 2, pp. 231–250, 2006.

[21] V. Lippi, T. Mergner, and G. Hettich, “A bio-inspired modular system for humanoid posture control,” in Proc. IROS Workshop Neuroscience Robotics “Towards robot-enabled, neuroscience-guided healthy society”, 2013, pp. 16–21.[22] G. Hettich, L. Assländer, A. Gollhofer, and T. Mergner, “Human hip ankle coordination emerging from multisensory feedback control,” Hum. Mov. Sci., vol. 37, pp. 123–146, Aug. 2014.[23] A. Alexandrov, A. A. Frolov, and J. Massion, “Biomechanical analysis of movement strategies in human forward trunkbending. I. modeling,” Biol. Cy-bern., vol. 84, no. 6, pp. 425–434, 2001.[24] A. D. Kuo and F. E. Zajac, “Human standing posture: Multi-joint move-ment strategies based on biomechanical constraints,” Prog. Brain Res., vol. 97, pp. 349–358, ISBN: 978-0-444-81252-0, Chapter 31 1993.[25] B. Henze, A. Werner, M. A. Roa, G. Garofalo, J. Englsberger, and C. Ott, “Control applications of TORO—A torque controlled humanoid robot,” in Proc. IEEE-RAS Int. Conf. Humanoid Robots, 2014, p. 841.[26] M. H. Woollacott, A. Shumway-Cook, and L. M. Nashner, “Aging and posture control: Changes in sensory organization and muscular coordina-tion,” Int. J. Aging Hum. Develop., vol. 23, no. 2, pp. 97–114, 1986.[27] M. Cenciarini, P. J. Loughlin, P. J. Sparto, and M. S. Redfern, “Stiffness and damping in postural control increase with age,” IEEE Trans. bio-med. Eng., vol. 57, no. 2, pp. 267–275, Jan. 2010.[28] F. Antritter, F. Scholz, G. Hettich, and T. Mergner, “Stability analysis of human stance control from the system theoretic point of view,” in Proc. Euro-pean Control Conf., 2014, pp. 1849–1855.[29] A. D. Kuo, “An optimal state estimation model of sensory integration in human postural balance,” J. Neural Eng., vol. 2, no. 3, p. S235, 2005.

Christian Ott, the Institute of Robotics and Mechatronics, German Aerospace Center, Wessling, Germany. E-mail: [email protected].

Bernd Henze, the Institute of Robotics and Mechatronics, Ger-man Aerospace Center, Wessling, Germany. E-mail: [email protected].

Georg Hettich, the University Hospital Freiburg, Freiburg, Ger-many. E-mail: [email protected].

Tim Niklas Seyde, the Institute of Robotics and Mechatronics, German Aerospace Center, Wessling, Germany. E-mails: [email protected], [email protected].

Maximo A. Roa, the Institute of Robotics and Mechatronics, German Aerospace Center, Wessling, Germany. E-mail: [email protected].

Vittorio Lippi, the University Hospital Freiburg, Freiburg, Ger-many. E-mail: [email protected].

Thomas Mergner, the University Hospital Freiburg, Freiburg, Germany. E-mail: [email protected].

Good Posture, Good Balance · idea is generalized to attain balance and posture control for a...

Documents

Transcript of Good Posture, Good Balance · idea is generalized to attain balance and posture control for a...