Control of Soft Pneumatic Finger-like Actuators for Affective MotionGeneration*

Mohammadreza Memarian1, Rob Gorbet2, Dana Kulic1

Abstract— This paper investigates the design and imple-mentation of a finger-like robotic structure capable of re-producing human hand gestural movements performed by amulti-fingered, hand-like structure. In this work, we present apneumatic circuit and a closed-loop controller for a finger-likesoft pneumatic actuator. Experimental results demonstrate theperformance of the pneumatic and control systems of the softpneumatic actuator, and its ability to track human movementtrajectories with affective content.

I. INTRODUCTION

User perceptual studies indicate that human hands canbe used for a variety of communicative purposes, includinggestural and affective communication [1] . Previous workshave found that velocity and acceleration of the motionstrongly influence perception [1]. In this paper, we proposea pneumatic circuit and closed loop feedback control designof soft pneumatic artificial muscle (SPAM) actuators capableof providing the required position, velocity and accelerationprofiles to convey gestural and affective movement.

Pneumatic artificial muscles (PAMs) are a group of pneu-matic actuators that convert the potential energy stored inpressurized gas to kinetic energy using an elastic membrane.The McKibben pneumatic muscle was one of the first PAMdesigns [2]; a variety of designs have appeared since then[3]. In general, PAMs consist of two main components:an elastic membrane and a guiding mechanism. The elasticmembrane expands volumetrically when pressurized, whilethe guiding mechanism constrains this expansion towards adesired direction determined by the physical design. Mod-elling these actuators is a difficult task because of the non-linear relationship between the change in the pressure of theelastic membrane and the change in the actuator’s geometryand volume [4]. The non-linear response of these actuatorscomplicates their position or force control. Joupilla andEllman [5] and Minh et al. [6] use linear PID pressure andposition controllers in cascade mode and achieve acceptableposition tracking results. In their design, the inner loop ofthe cascade controller uses pressure feedback to control thepressure at the intake of the actuator. The outer loop uses thereference position trajectory and position feedback to providethe reference pressure for the inner-loop controller.

In the more conventional designs of PAM actuators such asthe McKibben muscle, the guiding mechanism constrains theactuator to only produce longitudinal motion. In recent years,

* This work was supported by NSERC through the Discovery Grantprogram.

1 Department of Electrical and Computer Engineering, University ofWaterloo, Canada

2 Department of Knowledge Integration, University of Waterloo, Canada

more complex guiding mechanism designs have allowed forthe production of PAMs that produce a variety of motiontrajectories [7]–[10]. In recent literature, these actuators areoften referred to as soft actuators or soft PAMs (SPAMs).SPAMs possess the benefits of general PAMs such as com-pliance, high power to weight ratio, inherent stability andhigh speeds of movement [3]. These advantages, along withthe ability to design custom trajectories with SPAMs and thefact that these actuators can deliver power with no need fora support structure, have enabled a variety of applications.Examples include: exoskeletons and rehabilitation assistance[11], [12]; handling and grasping objects with complexgeometry, or delicate objects such as food [7], [13]–[17]; andmulti-DOF, multi-actuator, robotic mechanisms capable ofperforming sophisticated motions such as gait or swimming[8], [10], [18].

There are several SPAM designs that produce finger-likemotion [7]–[9]. PneuFlex [7] is an actuator designed forgrasping applications. Its rest pose and its motion trajectoryboth resemble that of the human finger. In comparison toother finger-like SPAMs, PneuFlex maintains its finger-likeappearance during its motion, avoiding the distension oftenobserved in SPAM designs (e.g., [9]). PneuFlex SPAMs arealso simpler to construct than the SPAM design from [8]for example, which requires more complex elastic membraneand guiding mechanism designs.

In applications such as grasping [7], low speed gait [18],and rehabilitation assistance [11], open-loop feedforwardcontrollers are used to control the SPAMs. These applicationsare mainly exploiting the compliance properties of SPAMsto conform to the body or grasped object and provideappropriate force, at relatively slow velocities and acceler-ations. To the best of the authors’ knowledge, no work onSPAMs focuses on controlling the fast dynamic response ofthese actuators to generate high velocity and accelerationmotions. In particular, no work to date achieves closed-loop control of the high speed response of these actuators.In this paper, an electromechanical system with pneumaticand electronic circuitry is designed that enables closed-loop control of PneuFlex SPAM actuators. First, the existingPneuFlex design is modified in order to make it suitable forfast dynamic responses and improve its durability. Second, apneumatic circuit that provides variable pressure with lowlatency and hysteresis is designed that utilizes innovativelow-cost mechanisms to improve noise levels in the pressurecontrol loop, enabling the system to achieve fast dynamicfeedback control. Third, linear PID pressure and positioncontrollers in a parallel configuration are used to control the

2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)Congress Center HamburgSept 28 - Oct 2, 2015. Hamburg, Germany

978-1-4799-9993-4/15/$31.00 ©2015 IEEE 1691

position output of the SPAM. In order to achieve feedback onthe position response of the SPAM, the actuator is equippedwith a gyroscopic sensor that provides feedback on theorientation of the tip of the actuator.

In the following sections of the paper, first, the approachfor achieving the goals of the work are discussed. In thissection, the modifications to the current designs of PneuFlexactuators, the pneumatic circuit, and the design of the closedloop controller are described. In section III, the experimentsconducted to examine the performance of the closed-loopcontroller are discussed. Section III also describes how thereference signal for the controller is extracted from theaffective motion trajectories. Also, the requirements on theperformance of the controller and how well the controller iscapable of meeting them are discussed. Lastly, a conclusionis given on the achievements of the goals of the work.

II. PROPOSED APPROACH

A. Custom designed SPAM Actuator

SPAM actuators are capable of performing a variety ofmotion trajectories depending on the design of their elas-tic membrane and guiding mechanism. In this work, anSPAM design which produces trajectories similar to fingermovements is developed based on the PneuFlex SPAM[7]. PneuFlex SPAMs have a cuboid shape at rest and aresimilar in size to the human finger. These actuators possesstwo unique features that allow them to perform finger-likemotions. First, the unique threading of fabric around theSPAM forces it to move in a straight direction along thelong axis of the SPAM when pressurized. The threading alsoprevents overstretching of the SPAM which in many SPAMdesigns results in an undesirable distended look when theSPAM is pressurized. Second, a flexible and non-elastic meshis attached to one of the four long faces of the SPAM. As theSPAM woven with fabric is pressurized, without the mesh,it elongates in a straight line. The attachment of the meshincreases the stiffness of the SPAM along one face, forcingit to bend in a curved trajectory towards the face with theattached mesh.

The design of the PneuFlex provided by the Roboticsand Biology Laboratory at the Technische Universitat Berlinwas modified to increase the air supply intake and improvethe fabric threading. The air pressure intake of the originalPneuFlex is located on the side of the actuator and the intaketube is considerably narrower than the internal hollow coreof the SPAM. The narrower air intake affects the pressureresponse of the SPAM by reducing the mass flow rate anddampens its motion due to the low pass filtering effects ofthe narrow intake. As illustrated in figure 1, by bringingthe intake of the SPAM to the bottom of the actuator itbecame possible to increase the intake tube size to 1/4”tubing which is much closer in size to the width of the SPAMhollow core. During the early stages of experimentation,it was also observed that the PneuFlex actuators showedsigns of damage at the points where the fabric threads meetthe edges of the SPAM. In order to improve durability andprevent this damage, wide ribbons were used which do not

cut through the silicon as easily. The resulting custom builtSPAM structure and its curved motion trajectory is shown infigure 2. From this point on, curved SPAM (CSPAM) willbe used to refer to this custom built SPAM structure.

As shown in figure 2, the motion of the CSPAM followsa curved trajectory. This trajectory modifies the orientationof the end effector (the distal end of the CSPAM) withrespect to the fixed reference frame. This means that the endeffector orientation can be used as feedback on the curvatureof the CSPAM structure. An ITG3200 gyroscopic sensor(InvenSense) is placed at the end effector of the CSPAMin order to measure orientation by integrating the measuredangular velocity. This orientation feedback is used by theclosed loop controller to control the curvature of the CSPAMto perform various affective motions, as described in sectionII-C.

Nonelastic Element

1/4” Air Intake

Gyro/Accelerometer

Wide Ribbon Threading

Hollow Core

Fig. 1. Components of the custom designed silicone-based SPAM withcurving motion trajectory (CSPAM)

Ref. Frame

End E!.

Frame

Fig. 2. The motion frames of the custom built SPAM with curved motiontrajectory (CSPAM). (Still frames taken from video; more video availablein supplementary material.)

B. Pneumatic Circuit

To model CSPAM actuators, similar to conventional PAMs[4], conservation of energy can be used to determine therelationship between the internal pressure P, actuator volumeV , torque τ and angular position θ . The law of conservationof energy states that, dWout = dWin where for CSPAMsthe output work is dWout = −τdθ and the input work isdWin = (P−Patm)dV . Combining these equations results in:

−τdθ = (P−Patm)dV. (1)

Equation (1) suggests that the output torque and angulardisplacements of the CSPAMs are determined by its internal

1692

pressure and its changing geometry during motion. Sincethis geometry change is pre-determined by the physicaldesign, internal pressure is the only parameter available forcontrolling the force and position output of the CSPAM.In this work, only the position output of the CSPAM is aconcern and the force output is not analysed.

Pulse Width Modulation (PWM) of high-speed valves( [5], [19], [20]) is used to control the internal pressure ofthe CSPAM actuator in this work. The main benefits of PWMwith high-speed valves are low size and cost, quick responsetime, low hysteresis, and relatively linear operating regions[20]. In the PWM approach, the duty cycle of the high-speedvalve determines the output pressure of the valve. Thus, thisduty cycle is the control signal of the feedback closed-loopcontroller.

Figure 3 depicts the pneumatic circuit of the system. Thehigh-speed 3/2 valve used is Festo’s MHE2-MS1H-3/2G-M7-K, capable of switching at 330Hz. The fast switchingprovides control over the output pressure of the valve. Pres-sure is provided to the valve by a pressure source (P15 TC byWerther Internationals) and pressure regulator (NAW2000-N02-2 by SMC) pair. The CSPAM and the pressure sensor(MPX5500DP by Freescale) are located on the right topcorner of the circuit. The pressure sensor is used for feedbackof the internal pressure of the CSPAM. This sensor is coupledwith an electronic low pass filter for noise reduction; the timeconstant of the filter is 2ms.

One of the main disadvantages of using high-speed valvesfor pressure control is that the output pressure is not smooth.The output pressure of the valve is produced by the valvepushing in or letting out high pressure packets of air at therate of switching. This discrete application and exclusionof air, while providing variable mean pressure at differentswitching duty cycles, results in a high magnitude of noiseat the output pressure of the valve. This noise makes thereadings from the pressure and the gyroscope sensors noisyas well, which will complicate the feedback control taskusing these readings. In this work, in order to reduce themagnitude of noise on the pressure output of the PWMvalve, a pneumatic Low Pass Filter (LPF) is designed andimplemented. Figure 4 shows a picture of the pneumatic LPF.To aid in understanding the function of this pneumatic circuit,an electrical analog is shown in Figure 5. The LPF con-sists of two main pneumatic components: the pressure tank(analogous to C2 in Figure 5) and the pneumatic resistance(R3). The pressure tank is a syringe. Pneumatic resistanceis provided by inserting a porous plug into the output tubeof the LPF (here, we used a pipe cleaner). This low-costnoise filtering solution provides the ability to fine tune thelow pass filter in a simple and quick manner by changingthe volume of the syringe or the length of the porous plug.By tuning these two parameters, at 3mL syringe volume and2.5cm porous plug length, the magnitude of oscillation onpressure caused by PWM was reduced to less than ±5%at 50% duty cycle. The inclusion of the LPF dampens thesystem and reduces the maximum achievable velocities. Inorder to rectify this issue, for the angular velocity during

pressurization, i.e., during the closing motion, the pressureis increased until required velocities are reached. Duringdepressurization, i.e., opening motion, the CSPAM, whichis a capacitive and a resistive load, is depressurizing throughits own resistance (R1) and the resistance of the pneumaticlow pass filter (R3), through the open valve (analog, Q1).In order to decrease the time constant of this circuit, a pulldown pneumatic resistor (R2) is introduced at the pressureintake of the CSPAM that reduces the time constant of thedepressurization circuit by reducing its overall resistance.

The low cost and easily tunable pneumatic resistor andlow pass filter enable the PWM signalling of fast responsepneumatic valves in order to control the output pressure ofthe circuit. These components reduce noise to levels suitablefor feedback closed-loop control. The reduction is achievedwithout compromising the benefits of using PWM singalingof fast valves, which are: low latency, low hysteresis, andnear-linear operation region. The CSPAM is also capableof achieving fast dynamics with the inclusion of thesepneumatic components.

Pull down

Pneu. Resistance

CSPAM

LPF Pneu.

Resistance

LPF Tank

High-speed

3/2 Festo Valve

Pressure

Regulator

High Pressure

Source

Exhaust

P

Pressure

Sensor

PCSPAM

Pvalve

PReg

Patm

PS

Fig. 3. The pneumatic circuit for the control of the CSPAM actuator

PIn

Filtered POut

Porous plug

Pressure Tank

Fig. 4. Pneumatic low pass filter

C. Controller

Considering the plant of the CSPAM actuation systemincluding the pneumatic circuit, the system can be dividedinto two sub-plants, as depicted in figure 6. The first sub-plant represents the pneumatic circuit of the system and is

1693

VCSPAM

VValve

Fig. 5. Electrical circuit analogy of the pneumatic circuit with low pass filter and voltage divider. Note that the CSPAM is a non-linear capacitor as itchanges volume at different internal pressures.

mainly governed by the switching response of the valve,the LPF and the pull-down resistor, and the CSPAM as apneumatic load. The input to the plant is the duty cycleof the valve and the output is the pressure at the intakeof the CSPAM. The second sub-plant of the system isthe mechanical structure of the CSPAM actuator and itsend effector orientation response with respect to its internalpressure. The input of the plant is the pressure at the intakeof the CSPAM and the output of the plant is the end effectororientation. Feedback is provided for the output of both ofthese sub-plants using the pressure and gyroscope sensors.The sub-plants are connected in series and there is no directcontrol over the input of sub-plant 2, i.e., the pressure atthe intake of the CSPAM can only be controlled by thepneumatic circuitry. Thus the controller has to be capableof controlling the end effector orientation of the CSPAMusing the valve duty cycle as its control signal.

The controller depicted in figure 6 is similar to cascadecontrollers previously used for PAMs [5], [6]. However,instead of using the outer loop position controller to providea reference for the inner loop pressure controller, the pressurereference trajectory is obtained by assuming a linear relation-ship between the internal pressure change of the CSPAM andits end effector orientation. This modifies the cascade con-troller designs in [5], [6] to a parallel configuration of pres-sure and position controllers. The linear relationship betweenthe pressure and the orientation of the CSPAM is obtainedfrom the open-loop analysis of the pressure step responseof the actuator. At internal pressure of 250kPa the CSPAMreaches around 3.97 rad steady state orientation. Using thesevalues, the relationship between pressure and orientationat 101kPa atmospheric pressure is, p = 250−101

3.97 θre f + 101.This relationship is used to calculate the required pressurereference signal from the end effector orientation referencesignal. Both controllers have duty cycle signals as theiroutput which are summed in order to determine the finalcontrol signal. The duty cycle signal has 12-bits of resolutionand the frequency of the PWM signal is 300Hz. With respectto this duty cycle resolution, the PID gains of the positioncontroller are, 200, 10, and 20000 respectively. The gain forthe pressure controller was 1.63. The controller is tuned sothat the control signal is mainly determined by the pressurefeedback controller and the position controller will only finetune the resulting motion of the CSPAM actuator towards thedesired position trajectories. A Teensy 3.1 microcontroller

was used for the implementation of this controller. Thecontroller loop was set at 800Hz, which is the highestsampling rate the gyroscopic sensor supports.

By taking advantage of the modifications and solutionsdiscussed in the previous sections, such as reducing thenoise of PWM signalling of the fast pneumatic valve orproviding full measurements of the dynamics of the CSPAM,it is shown that a linear cascade controller is capable ofcontrolling the non-linear CSPAM actuator. The performanceof this controller is presented in the following section.

III. EXPERIMENTS

A. Reference Signal

The parallel controller depicted in figure 6 requires orien-tation reference trajectories. The goal of the controller is todrive the CSPAM actuator to emulate, as closely as possible,the opening and closing of a single finger during affectivehand movement. An existing finger joint angle dataset [21]of affective hand opening and closing movements was usedto generate the reference trajectories. Three emotions areconsidered in that work: sadness, anger, and joy. For themovements collected in the dataset, there is a high correlationbetween the rotation of the metacarpophalangeal (MP) joint,the proximal interphalangeal (PIP) joint, the distal inter-phalangeal (DIP) joint, and the orientation of the finger’send effector calculated using forward kinematics. Duringthe motion for joy, for example, the correlations betweenthe MP, PIP, and DIP joints of the middle finger, were allabove 0.97. The correlations between each joint and the endeffector orientation were all above 0.98. Similar results wereobtained for the motions of anger and sadness. Thus, theend effector orientation trajectory of the middle finger waschosen as the reference signal for the controller; a gyroscopicsensor can be used to measure the tip orientation of theCSPAMs directly. In the dataset from [21], the joint angledata was collected using a data glove at 84 samples/second.The calculated fingertip orientation from this dataset wasinterpolated to 800Hz, and low pass filtered at a cut offfrequency of around 84Hz which is well above the Nyquistfrequency of the sampled data.

B. Control System Requirements

To enable the desired emotion to be conveyed throughmovement, the velocity and acceleration of the motion are

1694

PressureDuty Cycle

Pressure Gain Controller

Valve, Pneumatic Circuit &CSPAM Pressure Plant

CSPAM Pressure to Orientation Plant

1End EffectorOrientation

1ReferenceOrientation

Orientation PID Controller

Orientaion to PressureLinear Model

Fig. 6. Pressure-orientation closed-loop controller system for CSPAM actuators

critical [22]–[25]. Therefore, achieving the required veloci-ties and accelerations are an important requirement for thecontroller and the CSPAM actuators. Analysis of the CSPAMactuators during the design of the system has shown thatthe required peak velocities and accelerations are achievablein open-loop. Closed-loop control is required to ensure thatboth the peak velocities and accelerations and the expressivemodulations of the velocity during movement are accuratelyreproduced. Tracking error during the entire trajectory inorientation, velocity, and acceleration are also importantif the CSPAM movement is to resemble the closing andopening motion of the finger.

In addition to the quantitative requirements, there are qual-itative requirements for the closed-loop controlled motion ofthe CSPAM actuator. No part of the motion should divertthe observer’s attention away from the main goal of emotionconveyance. Therefore, behaviours such as visible vibrationsare highly undesirable.

C. Controller Results

Figure 7 shows the end effector orientation, its referencetrajectory, and the tracking error of the “joy” motion trajec-tory. The motion starts by closing the finger with medium ve-locities and accelerations. After performing subtle back andforth motion near the closing pose, the finger opens with highvelocity and acceleration. Since a short delay in the responseof the controller does not impede emotion conveyance, theresponse of the system has been slightly time-shifted toalign it with the reference signal. The plot of the trackingerror illustrates that, except during the closing motion ofthe CSPAM, the tracking error is small. In particular, thetracking error is low when subtle movements are performednear the closed pose, where the velocities are lower andtracking error will be more noticeable. The joy trajectoriesare shown here to illustrate the performance of the controller.The performance of the controller was similar for the sadnessand anger trajectories. The movements generated for all threeemotions can be seen in the supplementary video; additionalresults and trajectory examples can be found in [26].

For each of the three emotions investigated in [21],multiple joint trajectories were performed by the humandemonstrator. Within these trajectories, the one with the max-imum velocity was chosen as the reference trajectory for thatemotion. The variation between these recorded trajectories

is used in order to examine the tracking performance of thecontroller. The maximum variation between the trajectoriesare 0.87rads for anger, 0.34rads for joy, and 0.27rads forsadness. Table I presents the maximum magnitude and themean of the orientation tracking error. The maximum errorsof sadness and anger are below the maximum variationvalues of the human trajectories. The maximum trackingerror, which occurs during the closing motion of the joytrajectory, is greater than the maximum trajectory variationfor joy. This spike in tracking error is examined in thevelocity response of the controller below.

10.5 11 11.5 12 12.5 13 13.5 14 14.5 15−0.5

0

0.5

1

1.5

2

2.5

3

3.5

4

Orie

ntat

ion

(rad

)

a. Reference and Experimental End Effector Orientation for the Joy Motion

Expe.Ref.

10.5 11 11.5 12 12.5 13 13.5 14 14.5 15−0.5

0

0.5

1

Time (s)

Orie

ntat

ion

Err

or (

rad)

b. Orientation Tracking Error

Error

Fig. 7. a. Reference and experimental end effector orientation for the joymotion b. Orientation tracking error

TABLE IEND EFFECTOR ORIENTATION, VELOCITY, AND ACCELERATION

TRACKING ERROR

Joy Anger SadnessOri. Max Error (rad) 0.93 0.322 0.693Ori. Error Mean (rad) 0.00266 0.0133 0.0289Vel. Error Mean (rad/s) 0.0343 0.000398 0.00265Accel. Error Mean (rad/s2) 0.00689 -0.00328 0.0104

Figure 8 shows the end effector generated and referenceangular velocity, and the corresponding tracking error for

1695

the motion conveying the joy emotion. The angular velocityis calculated from the measured orientation using discretedifferentiation in Simulink. From the velocity plot, it can beobserved that the closing motion has high velocity trackingerrors as well. This error is mainly related to the differentvelocity profiles of the CSPAM actuator and the humanfinger (figure 7). As the velocity plot confirms, to reachthe same orientation value the finger accelerates quicklyand decelerates slowly while the CSPAM accelerates slowlyand decelerates quickly. This discrepancy in acceleration anddeceleration behaviours results in tracking error during theclosing motion. The velocity plot also shows that during theopening part of the motion, velocity tracking error is small.Oscillation of the end effector right after the opening motioncan be observed in the plot. This oscillation is undesirable asit can interfere with the affective perception of the motion.However, the orientation plot in figure 7 illustrates thatthe magnitude of oscillation in orientation is small. In thesupplementary video for the joy motion, it can be observedthat the oscillation is not noticeable.

Table I also presents the mean of the angular velocityand acceleration tracking error of all three emotions. Theacceleration of the motion is calculated using Simulink aswell. For both velocity and acceleration tracking, the meantracking error is small.

10.5 11 11.5 12 12.5 13 13.5 14 14.5 15−30

−25

−20

−15

−10

−5

0

5

10

Vel

ocity

(ra

d/s)

a. Reference and Experimental End Effector Angular Velocity for the Joy Motion

Expe.Ref.

10.5 11 11.5 12 12.5 13 13.5 14 14.5 15−10

−5

0

5

10

Time (s)

Vel

ocity

Err

or (

rad/

s)

b. Angular Velocity Tracking Error

Fig. 8. a. Reference and experimental end effector angular velocity forthe joy motion b. Angular velocity tracking error

TABLE IIEND EFFECTOR MAXIMUM AND MINIMUM VELOCITY AND

ACCELERATION ANALYSIS

Reference Achieved ErrorMax Velocity (rad/s) 31.5 28.6 2.89Min Velocity (rad/s) -28.2 -28.3 -0.0338Max Acceleration (rad/s2) 551 607 -55.2Min Acceleration (rad/s2) -664 -757 -92.4

Maximum and minimum required and achieved velocities

13.4 13.6 13.8 14−30

−25

−20

−15

−10

−5

0

5

Time (s)

Vel

ocity

(ra

d/s)

a. Joy critical opening action velocity

13.4 13.6 13.8 14−1000

−500

0

500

1000

Time (s)

Acc

eler

atio

n (r

ad/s

2 )

b. Joy critical opening action acceleration

10 10.2 10.4 10.6−10

0

10

20

30

40

Time (s)

Vel

ocity

(ra

d/s)

c. Anger critical closing action velocity

Expe.Ref.

10 10.2 10.4 10.6−1000

−500

0

500

1000

Time (s)

Acc

eler

atio

n (r

ad/s

2 )

d. Anger critical closing action acceleration

Fig. 9. Reference vs experimental motion response of the sections thatinclude maximum and minimum velocities and accelerations

and accelerations are presented in table II. The requirementson these motion parameters are determined by finding themaximum and the minimum peaks of the reference trajec-tories for all three emotions. The reference velocity andacceleration trajectories are calculated from the reference endeffector orientation signals using Simulink. The maximumrequired velocity and acceleration occur during the closingmotion of the angry trajectories. The minimum velocityand acceleration occur during the opening motion of thejoy trajectories. In other words, for the motions collectedin the dataset from [21], the angry motion requires themaximum closing motion velocity and acceleration, and thejoy motion requires the maximum opening motion velocityand acceleration. The plots in figure 9 present the referenceand response trajectories of velocity and acceleration duringthese critical sections of the reference signals. The resultsillustrate that except for the maximum velocity for the angrymotion, the system exceeds the minimum and maximumrequirements. Achieving the maximum velocity for the angrymotion was possible at higher PID gains, however, the higherPID gains produced undesirable overshoot which was moredistracting than the small error in maximum velocity.

The experimental results suggest that tracking of thedesired affective trajectories using the proposed cascadecontroller was achievable with small errors. These resultsalso show that the closed loop controller is capable of takingadvantage of the fast dynamics of the CSPAM actuators andcan enable them to achieve the high velocities and acceler-ations that were obtained during the open-loop analysis ofthese actuators. Qualitatively, the controller had acceptableresults as well. There were no undesirable motions suchas visible vibrations that would interfere with the affectivecapabilities of the motion.

IV. CONCLUSIONS

In conclusion, the closed-loop position controller is ca-pable of controlling the fast dynamic response of the

1696

CSPAM actuator and enables it to perform affective finger-like motions which require high velocities and accelerations.Our experimental results suggest that the cascade pressure-orientation controller allows the CSPAM actuator to achievethe performance specifications required for affective move-ments. Maximum velocities and accelerations were achievedwithin an acceptable error range in both directions of mo-tion. Mean tracking errors were at acceptable levels fororientation, velocity, and acceleration. From a qualitativeperspective, undesirable responses such as oscillation andtracking error were minimized, and were not observable inthe videos of the motion.

For applications that require more precise tracking, furtherreducing the pressure noise in the pneumatic circuit is oneof the main goals of our future work. The controller can alsobe improved by moving to more sophisticated linear or non-linear control strategies with model-based controllers whichtake the dynamics of the pneumatic system into account.Also, user studies to investigate the ability of the pneumaticactuator to convey affective gestures as part of a multi-finger,hand-like structure will be conducted.

ACKNOWLEDGMENT

The authors would like to thank the Robotics and BiologyLaboratory from Technische Universitat Berlin for providingus with their novel PneuFlex pneumatic actuators.

REFERENCES

[1] M. Karg, A.-A. Samadani, R. Gorbet, K. Kuhnlenz, J. Hoey, andD. Kulic, “Body movements for affective expression: a survey ofautomatic recognition and generation,” Affective Computing, IEEETransactions on, vol. 4, no. 4, pp. 341–359, 2013.

[2] H. A. Baldwin, “Realizable models of muscle function,” in Biome-chanics, pp. 139–147, Springer, 1995.

[3] F. Daerden and D. Lefeber, “Pneumatic artificial muscles: actuatorsfor robotics and automation,” European journal of mechanical andenvironmental engineering, vol. 47, no. 1, pp. 11–21, 2002.

[4] X. Shen, “Nonlinear model-based control of pneumatic artificialmuscle servo systems,” Control Engineering Practice, vol. 18, no. 3,pp. 311 – 317, 2010.

[5] V. Jouppila and A. Ellman, “Position control of pwm-actuatedpneumatic muscle actuator system,” ASME Conference Proceedings,vol. 2011, no. 54938, pp. 393–404, 2011.

[6] T. V. Minh, T. Tjahjowidodo, H. Ramon, and H. Van Brussel, “Cascadeposition control of a single pneumatic artificial muscle–mass systemwith hysteresis compensation,” Mechatronics, vol. 20, no. 3, pp. 402–414, 2010.

[7] R. Deimel and O. Brock, “A compliant hand based on a novelpneumatic actuator,” in Robotics and Automation (ICRA), 2013 IEEEInternational Conference on, pp. 2047–2053, IEEE, 2013.

[8] K. Suzumori, S. Endo, T. Kanda, N. Kato, and H. Suzuki, “A bendingpneumatic rubber actuator realizing soft-bodied manta swimmingrobot,” in Robotics and Automation, 2007 IEEE International Con-ference on, pp. 4975–4980, IEEE, 2007.

[9] Y. Sun, Y. S. Song, and J. Paik, “Characterization of silicone rubberbased soft pneumatic actuators,” in Intelligent Robots and Systems(IROS), 2013 IEEE/RSJ International Conference on, pp. 4446–4453,Ieee, 2013.

[10] F. Ilievski, A. D. Mazzeo, R. F. Shepherd, X. Chen, and G. M. White-sides, “Soft robotics for chemists,” Angewandte Chemie, vol. 123,no. 8, pp. 1930–1935, 2011.

[11] Y. S. Song, Y. Sun, R. Van Den Brand, J. Von Zitzewitz, S. Micera,G. Courtine, and J. Paik, “Soft robot for gait rehabilitation ofspinalized rodents,” in Intelligent Robots and Systems (IROS), 2013IEEE/RSJ International Conference on, pp. 971–976, Ieee, 2013.

[12] P. Polygerinos, Z. Wang, K. C. Galloway, R. J. Wood, and C. J.Walsh, “Soft robotic glove for combined assistance and at-homerehabilitation,” Robotics and Autonomous Systems, 2014.

[13] G. Fantoni, M. Santochi, G. Dini, K. Tracht, B. Scholz-Reiter,J. Fleischer, T. K. Lien, G. Seliger, G. Reinhart, J. Franke, et al.,“Grasping devices and methods in automated production processes,”CIRP Annals-Manufacturing Technology, vol. 63, no. 2, pp. 679–701,2014.

[14] C. Blanes, M. Mellado, and P. Beltran, “Novel additive manufacturingpneumatic actuators and mechanisms for food handling grippers,” inActuators, vol. 3, pp. 205–225, Multidisciplinary Digital PublishingInstitute, 2014.

[15] C. Eppner and O. Brock, “Grasping unknown objects by exploitingshape adaptability and environmental constraints,” in Intelligent Robotsand Systems (IROS), 2013 IEEE/RSJ International Conference on,pp. 4000–4006, IEEE, 2013.

[16] M. E. Giannaccini, I. Georgilas, I. Horsfield, B. Peiris, A. Lenz,A. G. Pipe, and S. Dogramadzi, “A variable compliance, soft gripper,”Autonomous Robots, vol. 36, no. 1-2, pp. 93–107, 2014.

[17] K. Suzumori, S. Iikura, and H. Tanaka, “Applying a flexible microac-tuator to robotic mechanisms,” Control Systems, IEEE, vol. 12, no. 1,pp. 21–27, 1992.

[18] R. F. Shepherd, F. Ilievski, W. Choi, S. A. Morin, A. A. Stokes, A. D.Mazzeo, X. Chen, M. Wang, and G. M. Whitesides, “Multigait softrobot,” Proceedings of the National Academy of Sciences, vol. 108,no. 51, pp. 20400–20403, 2011.

[19] R. B. Van Varseveld and G. M. Bone, “Accurate position controlof a pneumatic actuator using on/off solenoid valves,” Mechatronics,IEEE/ASME Transactions on, vol. 2, no. 3, pp. 195–204, 1997.

[20] M. Taghizadeh, A. Ghaffari, and F. Najafi, “Modeling and identifi-cation of a solenoid valve for pwm control applications,” ComptesRendus Mecanique, vol. 337, no. 3, pp. 131–140, 2009.

[21] A.-A. Samadani, E. Kubica, R. Gorbet, and D. Kulic, “Perceptionand generation of affective hand movements,” International Journalof Social Robotics, vol. 5, no. 1, pp. 35–51, 2013.

[22] F. E. Pollick, H. M. Paterson, A. Bruderlin, and A. J. Sanford,“Perceiving affect from arm movement,” Cognition, vol. 82, no. 2,pp. B51–B61, 2001.

[23] M. S. ADA, K. Suda, and M. Ishii, “Expression of emotions indance: Relation between arm movement characteristics and emotion,”Perceptual and motor skills, vol. 97, no. 3, pp. 697–708, 2003.

[24] J.-H. Lee, J.-Y. Park, and T.-J. Nam, “Emotional interaction throughphysical movement,” in Human-Computer Interaction. HCI IntelligentMultimodal Interaction Environments, pp. 401–410, Springer, 2007.

[25] M. Saerbeck and C. Bartneck, “Perception of affect elicited by robotmotion,” in Proceedings of the 5th ACM/IEEE international conferenceon Human-robot interaction, pp. 53–60, IEEE Press, 2010.

[26] M. Memarian, “Novel finger-like soft pneumatic actuators for affectivemovement,” M.S. thesis, ECE, UW, Waterloo, ON, 2015.

1697