FPGA-based Embedded System Designed for the Deployment in the

Compliant Robotic Leg CARL

Steffen Sch

utz

, Atabak Nejadfard

, Max Reichardt and Karsten Berns

Robotics Research Lab, Department of Computer Science, TU Kaiserslautern, Kaiserslautern, Germany

Keywords:

Embedded Systems, Distributed Control, Compliant Actuation, Series Elastic Actuator, Bipedal Walking.

Abstract:

The embedded system that is distributed within a bipedal robot is a key component of such a highly inter-

woven mechatronic system. Generally, it has to handle two competing main tasks – executing the embedded

closed-loop control of the actuators and handling the communication with the higher-level control system. As

the restrictions on physical size and energy consumption limit its computational resources, the design of the

embedded nodes poses a potential bottleneck for the performance of the overall system. Hence, the following

presents an approach to mitigate the conﬂicting requirements by deploying FPGA-based embedded nodes. It

is illustrated how the additional ﬂexibility at the logic level is used to implement the closed-loop force and

impedance control of a series elastic actuator. Furthermore, it is shown how the consequent hardware/software

co-design enables the deployment of a full featured robotic framework. To validate the concept, the properties

of the implementation are characterized.

1 INTRODUCTION

In an advanced bipedal robot, the distributed embed-

ded controllers should abstract the mechatronic sys-

tem for the higher-level control system – e.g., to act

like an ideal torque or position source (Radford et al.,

2015). This abstraction entails two main tasks at

the embedded level – the application speciﬁc tasks

such as executing the closed-loop control of the ac-

tuators and the handling of the communication with

the higher-level control. Depending on how much of

the closed-loop structure is distributed to the embed-

ded nodes determines how the criticality is distributed

between the communication and the application task.

If most of the closed-loop control cascades are exe-

cuted at the embedded nodes, the communication side

becomes less performance critical. In this scenario, a

high sampling rate and the deterministic execution of

the embedded control are the main determinant for the

achievable performance (Whitney, 1977; Shirai et al.,

2016). On the other hand, if a portion of the closed-

loop actuator structure is executed at the high-level

controller, the frequency and latency of the commu-

nication might limit the actuator performance (Zhao

https://orcid.org/0000-0001-7568-6439

https://orcid.org/0000-0001-9074-7519

https://orcid.org/0000-0002-9080-1404

et al., 2015). Thus, a deterministic execution with a

high frequency of the two main tasks is advantageous

for the performance of the overall biped.

Looking at the literature on the embedded solu-

tions deployed in current walking robots, it can be

noted that the degree of the distribution varies among

the systems. Table 1 summarizes the implementation

details as well as the main performance indicators

in form of various execution frequencies of several

legged machines. In all systems, with the exception

of the BioBiped, a dedicated controller handles the

current control of the electromagnetic motors. Thus,

the current control and – in case of BLDC motors the

commutation – is not mentioned explicitly in Table 1.

The embedded systems developed for Star-

lETH (Hutter, 2013) and BioBiped (Scholz, 2016) are

representatives of a centralized approach. As men-

tioned above, in case of the BioBiped, even the cur-

rent control is handled by the central high-level con-

troller. In StarlETH, the current and velocity con-

trol of the motors, are handled by the embedded con-

troller. Notably, the purpose of the latter is the lo-

cal and thereby high-frequent compensation of the

undesired motor and gearbox dynamics. Neverthe-

less, in both cases the actuator behavior desired by

the higher control layers are established by the high-

level controller. All other implementations that are

listed, namely DLR’s TORO, NASA’s Valkyrie, IIT’s

Schütz, S., Nejadfard, A., Reichardt, M. and Berns, K.

FPGA-based Embedded System Designed for the Deployment in the Compliant Robotic Leg CARL.

DOI: 10.5220/0007978205370543

In Proceedings of the 16th International Conference on Informatics in Control, Automation and Robotics (ICINCO 2019), pages 537-543

ISBN: 978-989-758-380-3

537

Table 1: Overview of the control infrastructure deployed in existing walking machines presented in the literature.

Platform C/D EC f

Bus f

Bus

HLC f

HLC

BioBiped (Scholz, 2016) C - - EtherCAT 1kHz Orocos/ROS 1 kHz

StarlETH (Hutter, 2013) C V 1 kHz CAN 400 Hz SL 400 Hz

TORO (Ott et al., 2010) D P/I/T 3 kHz Sercos-II 1 kHz Matlab 1 kHz

Valkyrie (Radford et al., 2015) D I/T 5 kHz Robonet - ROS < 3 kHz

WALK-MAN (Tsagarakis et al., 2017) D P/T 1 kHz EtherCAT 500Hz YARP 500Hz

ESCHER (Knabe et al., 2015) D I/T 2 kHz CAN 500 Hz Bitfrost/ROS 500 Hz

C – Centralized, D – Decentralized, EC – Embedded Control, HLC – High-Level Control

P – Position Control, V – Velocity Control, I – Impedance Control, T – Torque Control

WALK-MAN, and Virgnia Tech’s THOR/Escher –

implement at least the torque/force plus one further

control cascade at the embedded node.

Generally, it can be observed that the sampling

rate of closed-loop cascades that establish the ﬁnal ab-

straction towards the higher control layers is limited

in centralized control approaches – mostly due to the

bandwidth of the communication bus (Hutter, 2013).

Furthermore, as mentioned above, the latency intro-

duced by the communication bus reduces the stabil-

ity margin of the control system. Nevertheless, even

in highly distributed control approaches, a high band-

width communication with the superimposed control

layers is desirable.

In addition to the pure performance related char-

acteristics of an embedded system, another aspect is

the ease of handling of such a complex mechatronic

system as a bipedal robot. Usually, there is a break in

the development process between the powerful PCs

that execute the high-level control and the distributed

embedded nodes that implement the actuator control.

As Table 1 shows the high-level control is usually im-

plemented using a framework that runs on top of a

full-featured OS. The embedded nodes in contrast are

mostly used bare-metal or at most with a slim real-

time OS. Nevertheless, a more uniﬁed approach to

software development and especially system debug-

ging is desirable and facilitates the system handling

tremendously.

Therefore, in prior work, we presented an ap-

proach for the decoupling of the two competing tasks

by using an FPGA-based embedded node that enabled

the deployment of the full-featured robotic framework

Finroc

to the bare-metal embedded nodes (Sch

utz

et al., 2014; Reichardt et al., 2017). Following a

strict HW/SW co-design approach, the ﬂexibility pro-

vided by FPGAs ensured this did not negatively im-

pact the performance characteristics of the embedded

nodes. The deployment of either task-speciﬁc pro-

cessing units or general-purpose soft-processors al-

https://www.ﬁnroc.org/

lows for parallel computation to mitigate computa-

tional bottlenecks.

The concept for the FPGA system is illustrated

in Figure 1. At the center of each subsystem –

the communication side and the application side –

is a soft processor. The communication between

the two subsystems is handled via a block of Dual-

Port Ram. More details about the Ethernet-based

framework-integrated embedded protocol (FinEmbP)

and the data handling within the FPGA system are

provided in (Sch

utz et al., 2014).

This system served as the basis for the devel-

opment of the embedded nodes that are deployed

within the compliant robotic leg Compliant Robotic

Leg (CARL) developed at the Robotics Research Lab

(RRLab) (Sch

utz et al., 2017). It poses a ﬁrst it-

eration, to bring the biologically-inspired behavior-

based bipedal locomotion control (B4LC)– also de-

veloped at the RRLab – from a simulation to a phys-

ical robotic system. Thus, CARL is a bio-inspired

system that – inspired by the morphology of humans

– integrates mono- as well as biarticular actuations.

The redundant drive system – ﬁve series elastic ac-

tuators (SEAs) that act on three joints – is imple-

mented using two scaled RRLab SEAs developed at

the RRLab (Sch

utz et al., 2016).

Hence, the following is meant to give an insight

into the conceptual design and implementation of the

embedded nodes at the electronic as well as the FPGA

level. The purpose is to encapsulate the RRLab SEAs

towards B4LC as force and impedance sources while

exploiting the full actuator potential. Especially the

extensions to the HW/SW system at the application

side of the FPGA and the distribution of the closed-

loop tasks within that system are detailed.

2 ELECTRONIC SYSTEM

As a ﬁrst design iteration, at the electronics level, a

modular approach has been adopted. Therefore, as

illustrated in Figure 2a, the functionalities are dis-

ICINCO 2019 - 16th International Conference on Informatics in Control, Automation and Robotics

538

FPGA

Communication

Processor

Dual-Port

RAM

Application

Processor

IP Cores

MAC

Ethernet

Figure 1: Task decomposition within the FPGA system. Adapted from (Sch

utz et al., 2014).

tributed across several modules in form of separate

PCBs. One module contains the FPGA and the nec-

essary periphery. At the core of that module is the

eSM Board from elrest Automationssysteme GmbH

(for more information refer to (Sch

utz et al., 2014)).

The second module contains the output stage that is

used for the commutation and the current control of

the BLDC motor. Currently, this is handled by the

Gold Twitter servo drive from Elmo Motion Control

Furthermore, that module contains an ADC to capture

the winding temperature via the PT1000 temperature

sensor. A third module implements the I/O interface

to the embedded sensory system. For electric stability

and robustness, all signals are differential driven us-

ing RS-485 transceiver. The connection between the

modules is established using a backplane. Figures 2b

and 2c shows the modular backplane system.

In a second iteration, once the exact requirements

are known and the functionality of the electronics is

ensured, the modular system can be integrated into

a compact dedicated embedded node. Especially the

commutation and the current control of the BLDC

motor can also be handled by the FPGA making the

Gold Twitter obsolete.

3 APPLICATION-SIDE HW/SW

IMPLEMENTATION

Figure 3 shows a detailed illustration of the FPGA

system that implements the application side. At the

heart of it is the application soft processor – in this

implementation a Intel Nios II

. It is extended by a

multitude of IP Cores that are accessible by the Nios

processor as memory-mapped slaves. They handle

the serial interfaces to the embedded iC-MU-based

sensors. One IP Core directly interfaces the sensor

that captures the spring deﬂection which serves as the

http://http://www.elrest-gmbh.com/

http://www.elmomc.com/

https://www.intel.com/content/www/us/en/products/

programmable/processor/nios-ii.html

feedback for the force control of the SEA. Through

the dedicated IP Core, the sensory information can be

obtained at the maximum sampling rate and subse-

quently ﬁltered without causing any overhead at the

Nios. The ﬁltered spring position can be obtained by

a single memory access.

The sensor that captures the position of the rotor is

interfaced by the Elmo Gold Twitter as it requires the

information for the commutation of the electromag-

netic ﬁeld. Hence an IP Core within the FPGA taps

the SSI in which the Gold Twitter acts as the master.

Based on the obtained information another IP Core

calculates the absolute length of the ball screw. With

the rotational iC-MU being a single turn encoder, this

core detects the sensor overﬂows and hence – once

initialized – can calculate the absolute length of the

ball screw. After being ﬁltered in HW, the ball screw

length can be summed with the spring deﬂection to

obtain the actuator length that forms the feedback for

the impedance control.

The interface to the Elmo Twitter is implemented

by two PWM signals and a single I/O. Due to the

ﬂexibility of the FPGA, the properties of the PWM

signals could be exactly determined to fully exploit

the interface provided by the servo drive. Hence, the

PWM signal that encodes the desired current is modu-

lated with a base frequency of 160 MHz as this resem-

bles exactly the frequency at which the servo drive

logs the signal. The second PWM signal is the cur-

rent feedback. For safety reasons and in case of sys-

tem failure, a watchdog that monitors the Application

Nios can disable the servo drive via the I/O that trig-

gers its safe torque off (STO).

Another dedicated IP Core interfaces the serial in-

terface of the ADC that monitors the winding tem-

perature of the BLDC. This information is used in a

model of the actuators thermal dynamics that is im-

plemented in SW on the Application Nios. Further-

more, three of the ﬁve embedded nodes additionally

interface the sensors that are located at the joints.

Similar to the spring deﬂection, the sensor is sampled

at maximum frequency and subsequently ﬁltered by

dedicated IP Core.

FPGA-based Embedded System Designed for the Deployment in the Compliant Robotic Leg CARL

539

FPGA

Board

elrest eSM

ETH1 ETH2

Interface

Board

RS485 Transceiver

Output

Stage

Elmo

Gold Twitter

ADC

Backplane

(a)

(b) (c)

Figure 2: Embedded electronics developed for CARL– (a) the underlying modular concept and the implementation of (b) the

FPGA module and (c) the Elmo based output stage. All the three boards are mounted on the backplane board.

4 IMPLEMENTATION RESULTS

Figure 4 depicts the distribution of the closed-loop

cascades, the relevant data ﬂow, and the respective

sampling frequencies within the HW/SW system. The

current control and the commutation of the BLDC

motor are handled by Elmo Gold Twitter at a fre-

quency of 20 kHz. As mentioned above, this fre-

quency also dictates the sampling frequency of the

IP Cores that tap the rotor sensor, calculates the ball

screw length, and ﬁlter the respective data. The spring

deﬂection is sampled and ﬁltered at 40 kHz.

Similar to Hutter (Hutter, 2013), the force control

is established by a PID control. As already indicated

in Figure 3, the PD-portion of the force control is im-

plemented as a dedicated IP Core. It is executed at

a frequency of 10 kHz and, as it is implemented in

logic, without any jitter. The integral part of the force

control is implemented in software. The combination

of a low I-gain and an anti-windup mechanism is used

to eliminate the steady-state force error. The force

control is complemented by a feed-forward term that

is based on the motor torque constant.

The outer impedance cascade is also implemented

in software. It relies both on the spring deﬂection

and the ball screw length as feedback signals to de-

termine the current actuator length. All application-

side software is implemented within the deployed

Finroc instance and are executed with a frequency of

5 kHz. Notably, the application-side software imple-

ments more functionality beyond the closed-loop con-

trol. It includes the thermal model of the BLDC mo-

tor, several safety mechanisms, and the conversion of

the raw sensor values to the SI values that are passed

to the higher control layers.

Comparing the achieved sampling rates to the

numbers summarized in Table 1, it becomes obvious

that the presented system outperforms the comparable

systems by at least a factor of two.

ICINCO 2019 - 16th International Conference on Informatics in Control, Automation and Robotics

540

FPGA

Elmo

Twitter

A/B/C

PD Control

Current

Feedback

Application

Nios

Filter

SSI Master

iC-MU Spring

SSI Master

iC-MU Joint

Filter

SSI Listener

iC-MU Rotor

Length

Calculation

Filter

Watchdog

ADC

SPI Master

ADC Pt1000

PWM

SSI

STO

SSI

SPI

Figure 3: IP Cores acting as coprocessors on the application side of the FPGA system.

FPGA

Nios

5 kHz

Elmo

10 kHz

Current

Control

20 kHz

Impedance

Control

Filter

SSI Master

iC-MU Spring

40 kHz

Filter

Length

Calculation

SSI Listener

iC-MU Rotor

20 kHz

Force Control

Figure 4: Distribution of the closed-loop cascades within the HW/SW system with the respective sampling frequencies.

4.1 Bare-metal Finroc Execution

Timing

Besides the raw sampling rates, the performance of

the application-side software has been analyzed re-

garding the execution timing and jitter (Reichardt

et al., 2017). A dedicated IP Core has been used for

the measurement of the exact timing in clock cycles

of the application soft processor. Within the Finroc

execution cycle, the maximum jitter was around 569.4

cycles. This equals a variation of the execution timing

by 2.29 % relative the overall cycle length. Thus, the

value is well below the desirable 10 % that are deter-

mined by Shirai et al. as performance critical (Shirai

et al., 2016).

Overall, the numbers validate that no performance

limitations regarding the closed-loop control of the

actuator originate from the deployment of the robotic

framework to the embedded nodes. On the other side

it entails the beneﬁt that the nodes are seamlessly in-

tegrate into the tooling of the high-level control sys-

tem without a need to develop any additional tools.

Especially, it facilitates the task tremendously when

putting a complex system into operation.

FPGA-based Embedded System Designed for the Deployment in the Compliant Robotic Leg CARL

541

4.2 FinEmbP Timing

Although the presented system poses a highly dis-

tributed implementation of the actuator control that

reduces the requirements on the communication bus,

the timing of FinEmbP implementation is investigated

in a second experiment. The characteristics are eval-

uated using the FinEmbP network that is used for the

ﬁrst walking experiments with CARL. It consists of

seven FinEmbP nodes – four standard SEA nodes as

described above; one SEA node that additionally in-

terfaces the sensory system in the foot to capture the

interaction with the ground; a node that controls the

test-rig (TR) that consists of a treadmill and a winch

mechanism to mimic the second leg; one AD node

that monitors the supply voltage and current of the

overall system.

Table 2 gives the data that is transmitted and re-

ceived by the nodes. The table shows that the trans-

Table 2: Data sent and received by the FinEmbP nodes in

Byte [B].

SEA TR AD Total

Std Foot

Output 108 140 96 136 804

Input 56 56 4 60 344

Internal 332 368 108 328 -

Blobs 7 8 3 7

mitted data is relatively low – one of the beneﬁts of

the actuator abstraction towards the higher control

layers. Besides the explicit input/output data, node

internal data is divided into blobs with a maximum

size of 48 B that are attached to the cyclic output data

frames (referred to as iSend in the FinEmbP context,

refer to (Sch

utz et al., 2014)). Hence, this mechanism

poses a trade-off between the update rate of the inter-

nal data and the input/output data.

The seven nodes are arranged in a mixed star/line

topology as shown in Figure 5. It includes two

FinEmbP

Master

Switch

Node

Test-Rig

Node

CARL

Switch

SEA

Node

SEA

Node

SEA

Node

SEA + Foot

Node

SEA

Node

Figure 5: Topology of the FinEmbP network used during

the experiment.

switches – one that is located within CARL and the

second between CARL and the PC that executes the

FinEmbP master.

For the experiments, the FinEmbP master is run-

ning on a standard PC (Intel Core i7-4790K at 4 GHz

and 16 GB RAM) with a headless Ubuntu 14.04. The

high-level walking control developed for CARL– a

subset of B4LC– is executed on top of the FinEmbP

master. Hence, the overall setup resembles the normal

operating conditions. Notably, the FinEmbP protocol

is running at maximum frequency. The idle phase that

is part of the protocol is set to zero.

To obtain a representative impression of the prop-

erties, 50000 FinEmbP cycles have been recorded.

The average cycle time is 736 µs (approximately

1.36 kHz) with a standard deviation of 53 µs (7 % of

the average cycle time). The jitter is mostly due to

the implementation at the master side. As aforemen-

tioned, it is executed on a standard OS using a POSIX

UDP socket. Figure 6 shows the histogram of the

50000 cycles.

600 620 640 660 680 700 720 740 760 780 800 820 840 860

1,000

2,000

3,000

FinEmbP Cycle Time [µs]

Cycles

Figure 6: Histogram of the FinEmbP timing. The data is

recorded with a network of seven nodes and the high-level

walking control of CARL running on the master PC.

The average bus frequency compares well to the

comparable systems listed in Table 1. As the actuator

control is completely executed at the embedded node,

the observed jitter is not critical.

5 CONCLUSIONS AND FUTURE

WORK

The paper at hand presents application speciﬁc archi-

tecture of the dedicated embedded system designed

for the use in CARL, a compliant bio-inspired robotic

leg. Following a strict HW/SW-codesign approach,

the deployment of a full-featured robotic framework

to the bare-metal soft processor running at 100 MHz

is achieved. This is formulated as an explicit design

goal for the development of Robot Operating Soft-

ware (ROS) 2.0 (Gerkey, 2017).

Besides the soft beneﬁts that come with the ex-

tension of the robotic framework to the embedded

nodes, the performance critical properties are pre-

ICINCO 2019 - 16th International Conference on Informatics in Control, Automation and Robotics

542

served. This is achieved by implementing the sen-

sor sampling and ﬁltering as well as parts of the force

closed-loop control as dedicated IP Core. Especially,

the latter removes any signiﬁcant jitter from the con-

trol execution.

It has been validated experimentally that the exe-

cution characteristics of the software components im-

plemented within the robotic framework are adequate.

Overall, the achieved sampling frequencies and deter-

minism match or even exceed the properties of com-

parable systems. Similarly, the achieved frequency of

the FinEmbP communication bus with 1.36 kHz com-

pares well to other walking machines.

A potential next step would be the implementa-

tion of the commutation and the current control of

the BLDC within the FPGA fabric as a dedicated

co-processor. Thereby, the Elmo Gold Twitter servo

drive can be substituted by a bare output-stage that

features three half bridges and the appropriate facili-

ties for the current measurement. Subsequently, the

modular electronics can be composed to a highly-

integrated physically small single PCB. It could be

beneﬁcial to integrate the electronics into the actua-

tor.

If the jitter of the communication bus proofs to be

critical, ﬁrstly a real-time operating system (OS) at

the master side poses a potential improvement. Fur-

thermore, the OS’s stack could be partially bypassed

using, e.g., by using libpcap.

ACKNOWLEDGEMENTS

This work was partly funded by the European Com-

mission 7th Framework Program under the project

H2R (no.60069).

REFERENCES

Gerkey, B. (2017). Why ROS 2.0? http://design.ros2.org/

articles/why ros2.html. Accessed: 2019-04-16.

Hutter, M. (2013). StarlETH & co-design and control of

legged robots with compliant actuation. PhD thesis,

Diss., Eidgen"ossische Technische Hochschule

ETH Z"urich, Nr. 21073, 2013.

Knabe, C., Grifﬁn, R., Burton, J., Cantor-Cooke, G., Danta-

narayana, L., Day, G., Ebeling-Koning, O., Hahn, E.,

Hopkins, M., Neal, J., and others (2015). Designing

for Compliance: ESCHER, Team VALOR’s Compli-

ant Biped. Technical report, Terrestrial Robotics En-

gineering & Controls (TREC) Lab, Virginia Tech.

Ott, C., Baumg

artner, C., Mayr, J., Fuchs, M., Burge, R.,

Lee, D., Eiberger, O., Albu-Sch

affer, A., Grebenstein,

M., and Hirzinger, G. (2010). Development of a biped

robot with torque controlled joints. In Humanoid

Robots (Humanoids), 2010 10th IEEE-RAS Interna-

tional Conference on, pages 167–173. IEEE.

Radford, N. A., Strawser, P., Hambuchen, K., Mehling,

J. S., Verdeyen, W. K., Donnan, A. S., Holley, J.,

Sanchez, J., Nguyen, V., Bridgwater, L., et al. (2015).

Valkyrie: Nasa’s ﬁrst bipedal humanoid robot. Jour-

nal of Field Robotics, 32(3):397–419.

Reichardt, M., Sch

utz, S., and Berns, K. (2017). One

ﬁts more - on highly modular quality-driven design

of robotic frameworks and middleware. Journal

of Software Engineering for Robotics (JOSER):

Special Issue on Robotic Computing, 8(1):141–

153. ISSN: 2035-3928; this publication is available at

http://joser.unibg.it/index.php?journal=joser&page=is

sue&op=view&path[]=9.

Scholz, D. (2016). On the Design and Development of Mus-

culoskeletal Bipedal Robots. PhD thesis, Technische

Universit

at Darmstadt.

Sch

utz, S., Mianowski, K., K

otting, C., Nejadfard, A., Re-

ichardt, M., and Berns, K. (2016). RRLAB SEA – A

highly integrated compliant actuator with minimised

reﬂected inertia. In IEEE International Conference

on Advanced Intelligent Mechatronics (AIM).

Sch

utz, S., Nejadfard, A., Mianowski, K., Vonwirth, P., and

Berns, K. (2017). CARL – A compliant robotic leg

featuring mono- and biarticular actuation. In IEEE-

RAS International Conference on Humanoid Robots,

pages 289–296.

Sch

utz, S., Reichardt, M., Arndt, M., and Berns, K. (2014).

Seamless extension of a robot control framework to

bare metal embedded nodes. In Informatik 2014, Lec-

ture Notes in Informatics (LNI), pages 1307–1318,

Stuttgart, Germany.

Shirai, T., Osawa, K., Chishiro, H., Yamasaki, N., and

Inaba, M. (2016). Design and Implementation of

a High Power Robot Distributed Control System on

Dependable Responsive Multithreaded Processor (D-

RMTP). In IEEE 4th International Conference on

Cyber-Physical Systems, Networks, and Applications

(CPSNA), pages 19–24.

Tsagarakis, N. G., Caldwell, D. G., Negrello, F., Choi,

W., Baccelliere, L., Loc, V., Noorden, J., Muratore,

L., Margan, A., Cardellino, A., Natale, L., Hoffman,

E. M., Dallali, H., Kashiri, N., Malzahn, J., Lee, J.,

Kryczka, P., Kanoulas, D., Garabini, M., Catalano,

M., Ferrati, M., Varricchio, V., Pallottino, L., Pavan,

C., Bicchi, A., Settimi, A., Rocchi, A., and Ajoudani,

A. (2017). Walk-man: A high-performance humanoid

platform for realistic environments. Journal of Field

Robotics, 34(7):1225–1259.

Whitney, D. (1977). Force feedback control of manipulator

ﬁne motions. Trans. ASME J. of Dynam. Syst., Mea-

sur. and Control, 99(2):91–97.

Zhao, Y., Paine, N., Kim, K. S., and Sentis, L. (2015). Sta-

bility and Performance Limits of Latency-Prone Dis-

tributed Feedback Controllers. IEEE Transactions on

Industrial Electronics, 62(11):7151–7162.

FPGA-based Embedded System Designed for the Deployment in the Compliant Robotic Leg CARL

543