Science & Education

DRIBBEL 10 KG ROBOT WALKING ON 16 PENLITESir. Edwin Dertien, prof. dr. ir. Stefano Stramigioli,

University of Twente

At the Control Engineering group of the University of Twente a lot of research is spent on

various types of robots. This paper describes the work on the design and construction ofDribbel, a passive dynamic walking robot. Dribbel is a ‘2D-biped’: he can not fall sideways

because four legs in a row are used (comparable to a person walking on crutches). This

construction is chosen to focus on the forward motion, without having to solve the complex

3D stabilization.

The Power in Modelling

photos: courtesy of University of Twente

Dynamic walking means that the motion ofthe robot is induced mainly by the internaldynamics of the machine. Active energy inputby the actuators is only used to stabilize themotion or force the robot to increase itswalking speed. The international namingconvention for this type of research isDynamic Walking. Passive dynamic walkingmeans that a system must show a goodwalking from itself (passive) without feedingexternal energy, comparable to the motion ofa pendulum. Human walking is an example ofsuch a pendulum like motion.

The starting point for any dynamic walkingrobots is a simple pendulum. The goal is torealize a robust energy efficient walkingmotion. This is contrary to the well-knownJapanese walking robots such as Honda’sASIMO, where all degrees of freedom arecontrolled by conventional, energy wastingactuators.

This paper focuses on the practical aspects ofthe work on Dribbel: the design strategy, themechanical construction and the electronics.

Dynamic WalkingThe idea of using a set of pendulums tocreate a walking robot originates from thepatents on toys in 1888. Real research wasstarted by Tad McGeer [2] in the early 1990s.His constructions were able to walk down ashallow slope without any form of activecontrol or actuation. Base on these fullypassive ‘slope walkers’ a number of robotshave been built that do use actuators andcontrol. Dribbel is one of these robots.By using a motor in the hip, a slope is nolonger necessary. The robot can walk on alevel plane. The rest of the robot (knees andfeet) are still fully passive.

SimulationDuring the design of the robot, computersimulations have been extensively used. Firstsome simple models were created to find theoptimal weight distribution of the robot. Thesize of the robot was chosen roughly athuman size, with legs 1 m in length, a weightof roughly 10 kg primarily located in the hip.From this model, the required hip torque forthe robot was derived, requiring peaks of 10Nm.

Confident about the chosen components andsizes, a start was made designing the robot’smechanics in SolidWorks, while simulatingthe robots behavior in more detail in thesimulation environment.For the simulations, the software package 20-sim [5] was used. This package uses bond-graph notation (besides standard blockdiagrams and equations) in order to makepower-continuous domain-independentmodels. For the three-dimensional (3-D)kinematics and dynamics, the special 3-Dmechanics toolbox in 20-sim has been used,which provides the user with a simple drag-

photos: courtesy of Boskalis

Figure 1: The current design of the walking robotDribbel.

and-drop drawing interface for kinematicstructures (see Figure 2). The interface of thistoolbox is comparable to a 3D CADenvironment. Internally, this package derivesequations using screw theory [4]. In 20-simthe resulting kinematics and dynamics modelis coupled with the power interaction of themotors and the environment.

The resulting model was used for testing thecontroller algorithm, testing the effect ofadding extra battery weight, but also theeffects of various materials on the foot soles.After the mechanical prototype was built, thesimulation model has been fine tuned tomatch the exact robot behavior.

For future experiments a very accurate modelis available. Figure 3 shows the similaritybetween the hip angle in simulation andmeasurements during a test-run. Recently thesimulation model has been extended withactuated ankles. The actuated ankles arecurrently added to the real robot and wehope to do test runs in the fall of 2007.

ControlFor the hip actuator in both the simulationenvironment and the real robot, a simple PD-controller is used. The set point for thecontroller is switched on foot impact. Thissimple control has been used by other

powered “passive” designs [3]. By changingthe set point and controller gain, the walkinggait of the robot can be influenced. Thecontroller is tuned to have a very weakaction. The swing leg will reach the set pointbut will fall back immediately due to gravityso that the angle between the legs on impactis much smaller than the given set point. Atthe start of the swing phase, the controllercan be seen as passive spring connectedbetween the stance leg and swing leg.

The product of the set point and the gain is ameasure for the amount of initial torque withwhich the leg is being swung forward.

In order to create a pure passive mechanismwith an electric motor, the mechanism mustmove freely when the motor is turned off(back-driveability). Although the gearbox hasa high efficiency, some resistance ismeasured. Using a torque sensor at the rightplace and a proper control algorithm, thisresistance can be canceled out. Using this‘zero torque’ control algorithm, thecombination of motor and gearbox behavesas an additional mass, without damping themotion.

MechanicsThe hip is the most important joint in thisrobot, being the only one actuated. The kneesand ankles are passive joints. The hip is thus

Figure 2: Screenshot of the 3-D model made withthe 20-sim 3-D mechanics editor.

Figure 3: Hip angle during a short, straight walkin both the simulation and real robot.

designed around the main actuator: a MaxonRE40 150-W brushed DC motor connectedwith a gearbox with a large gear ratio (1:73). The mechanical design of the hip consists ofa 50-cm aluminum tube 6 cm in diameter.With large SKF bearings, an 8-cm-diametertube is fitted around this tube. The outer legsare mounted on the inner tube, the inner legson the outer tube.

The motor is mounted in the inner tube, andthe output power is transferred using anOldham coupling, via a torque sensor, to theouter tube as can be seen in Figure 4. Thetubular design was chosen because a tubehas the best known stiffness-to-weight ratiofor a hollow object. Moreover, it is a mountingspace for the motor, gearbox and batteries.

The legs consist of rectangular hollowaluminum bars that can be bolted ontocouplings to the hip tube, knees, and feet. Alljoints can be disconnected by simply removing

four screws, so the design is very modularand allows for easy installment of differentknees or feet modules.

So far two sets of feet have been tested: thefirst feet where small and flat with a thin layerof anti-slip rubber glued to the sole. The secondset had a thick layer of rubber. Simulationsshowed that the extra compliance of thesecond set results in a more efficient walkingmotion [7]. This was verified by mounting a setof bouncing balls under the feet.

For the knees, a locking mechanism had to bedesigned. Although the knees are notactuated during motion, they have to belocked, as soon a the weight of the robot iscarried.

The first design consisted of a locking systemwhere a solenoid had to retract a locking pin.This system failed terribly. When the leg wasunder stress, the solenoid could nevergenerate the amount of force necessary toretract the locking pin. A better solution wasthe use of door locking magnets. Although thissolution is not energy efficient (during walking

70% of the energy is used by the lockingmagnets) is turned out to work very well. Bothof the locking mechanisms can be seen inFigure 5.

Figure 4: SolidWorks drawing of the drive train.

Figure 5: The knee-lock mechanism with locking magnet and the mechanism with retracting pin.

ElectronicsThe tasks for the electronic system consist ofmeasuring, sensing and controlling the robot.A modulated distributed network is used onthe robot. This will prevent the use of loosecables and allow to add extra modules anddegrees of freedom after wards.

Each joint has its own controller board to readthe sensors (encoders and foot sensors) andsteer the actuators (knee locking magnets).The boards are connected by a TWI bus (two-wire interface, also known as Philips’ I2C bus).Therefore, only four wires (including powersupply) are needed to connect everything onthe robot. A frequency of 100 Hz is used tocollect all data in the robot for the maincontroller. The internal data rate of the boardis 40 kHz (see figure 7).

On the boards, Atmel ATmega8 RISC microcontrollers are used. These small microcontrollers are capable of nearly 16 MIPS at 16MHz. Hardware and interrupt services for theTWI bus are already implemented inside thiscontroller. The knee and feet encoders aremeasured with a relatively high frequency (40kHz). The maximum resolution for thestandard HP55xx series is 500 cpr (counts perrevolution). In quadrature, this results in 2,000cpr, yielding a resolution of 0.18º. Angularvelocities are calculated using an Eulerdifferentiation algorithm executed in the

micro controller. The controllers wereprogrammed using a propriarity C compilerfrom Codevision (http://hpinfotech.ro). The TWIroutines from the library from Procyon (http://hubbard.engr.scu.edu/avr/avrlib/) wereadapted for this compiler.

Each board contains four LED’s to indicatethe status of the board and an RS-232 portfor communication and fault detection. Themotor amplifier is connected to the TWI-bus,with the same ATmega8 micro controller asused for the knee and feet joints. The motoramplifier consists of a custom designed H-bridge using IR2110 half-bridge drivers.Safety-monitoring, temperature-sensing, andcurrent-limiting functions are performed bythe micro controller. A central relay can beused to turn off the power stage. Also, anautomatic fuse is added in the power stage.For noise-reduction spike-suppression diodes,capacitors and a small snubber network wereadded.

The hip motor uses the same HP5540-seriesencoder as used for the knees and feet. A PIDcontrol loop with an update frequency of 1 kHzis used to drive the motor. The set point andgain values are received over the TWI bus. Themost difficult part regarding the motoramplifier was designing a printed circuit board(PCB) that could be fitted inside the tube witha diameter of 6 cm. Especially the heat sink,relay, capacitors, and power regulators had to

Figure 6: The joint module located on the knee.This board interfaces the encoder, controls theknee-lock magnet, and is connected to the TWI

interface using the four colored wires.

Figure 7: A screen shot of the oscilloscopedisplaying TWI bus data.

be placed with care. After several attempts,even using cardboard mock-ups, a 5.5-cm-wideand 18-cm long PCB design was made,containing all components. Figure 8 shows thebreadboard design and the cardboard mockup.At the back of the motor a pack of 16 NiMhpenlites has been placed. This pack can yield avoltage of 19, V and has enough capacity to letthe robot run for more than 30 minutes.Besides the encoders for measuring allangles, a torque sensor is used in the hipjoint. For this sensor an amplifier board wasdesigned that is connected with the TWI-bus.The system can also be coupled to the localSPI-bus of the motor amplifier, to implementthe ‘zero torque’ control algorithm. For thetorque sensor, an interface containing aMAX1452 strain-gauge amplifier and againthe ATmega8 controller was designed. TheATmega on board fulfilled the tasks of TWIinterface and analog to digital (AD) converterand, using its serial interface, acted as aprogrammer for the MAX1452 IC.

The last micro controller circuit (bringing thetotal to a total of 11 micro controller boardsconnected to the same bus) is the ‘brain’ ofthe robot. It is used as central communicationprocessor and main walking algorithmcontroller. This board is equipped with anATmega128 running at 16 MHz. This

controller acts as the TWI bus master,gathering status data from all slaves (joints,torque sensor, motor amplifier) and sendingcommands to the knee locks and motoramplifier. The brain can send a full robotstate (all angles, switch status, powerconsumption, and torque) to a host PC with arate of 100 Hz for data-logging purposes.Data logging and analysis of the robot data isperformed in 20-sim.

ExperimentsThe robot has been walking around foralmost two years. Most of the walkingexperiments took place in a small laboratorywhere a stretch of 10 m can be used to letthe robot walk. Figure 9 shows a time-lapseshot of the robot while walking. For the firsttests, a safety line was used. The initial testboosted confidence in the stability and thesafety lines were removed.

A well-established criterion for the stability ofa walking robot is the distance between therobot and its designer during a test walk(attributed to Tad McGeer). Some studentsoverestimated the stability, causing the robotto fall on the ground before they could catchit. However, the robot survived all falls, with acouple of bent knee-caps and a brokenencoder-casing being the main damage.

Figure 8: A working breadboard design of the 150-W motor amplifier and a cardboard mock-up of thePCB design.

While being controlled by a single actuator,the robot can walk at various speeds, usingvarious types of walking. The robot is able towalk with a power consumption of only 1.6 W.This makes the robot a lot more efficientthan the Japanese ASIMO which uses around100 W. To make a fair comparison, theefficiency of robots is expressed as the costsof energy to carry a certain weight over acertain distance. For Dribbel, this cmt-value[1] is 0,06, which is very low, even whencompared with other passive dynamicwalking robots.

ConclusionThe design trajectory as described hereworked well. The parallel use of simulationand real-world testing yielded a workingprototype robot that is robust enough for thedaily lab experimental work. The matchedsimulation models proved valuable in testingnew controller algorithms and predicting thebehavior of new mechanical additions, such asthe compliant feet [7].

More InformationMore information on the building andconstruction of the robot can be found on theweb-site of Control Engineering(http://www.ce.utwente.nl). The progress ofthe walking robot can be monitored on theproject website at http://www.ce.utwente.nl/biped.

Acknowledgments The design and construction of Dribbel hasbeen a group effort: Niels Beekman, Gijs vanOort, Eddy Veltman, and Vincent Duindamcontributed heavily to this project undersupervision of Prof. Stefano Stramigioli. Thework on walking robot systems at theUniversity of Twente is performed at theIMPACT institute and located at the ControlEngineering group of the faculty of ElectricalEngineering.

References[1] S. H. Collins and A. Ruina, A bipedalwalking robot with efficient and human-likegait, 2005.[2] T. McGeer, Passive Dynamic Walking, Int. J.Robotics Res., vol. 9, par. 10.3, p 72, 1990.[3] M. Wisse and J. van Frankenhuyzen, Designand construction of Mike; a 2D autonomousbiped based on passive dynamic walking,AMAM, 2003.[4] S. Stramigioli and H. Bruyninckx, Geometryand screw theory for robotics, Tutorial duringICRA 2001, 2001.[5] Controllab products, 20-sim, ver3.6. http://www.20sim.com, 2006.[6] E.C.Dertien, Realisation of an energy-efficient walking robot, MSc thesis,022CE2005, June 2005.[7] E. Veltman, Foot shapes and ankle actuationfor a walking robot, MSc thesis, 002CE2006,May 2006.

Contact

info@20sim.com+31 (0)85 773 1872

© Controllab Products BV, Hengelosestraat 500 , 7521 AN Enschede, the Netherlands,

Figure 9: A time-lapse shot of the walking robot.