DLR High altitude balloon launched experimental glider(HABLEG): system design, control and flight data analysis

Maximilian Laiacker, Sven Wlach and Marc Schwarzbach

Abstract— In this paper the autopilot setup and thelessons learned from a successful balloon launched highaltitude glider mission will be presented. Flying anunmanned aerial vehicle inside the stratosphere meansoperating it outside visual line of sight, in low pressuresand cold temperatures. These environmental conditionspose many challenges to the UAV aerodynamic design,the autopilot system and the ground infrastructure. InMay 2015 the HABLEG airplane was carried up to analtitude of 20km by a balloon. When the target altitudewas reached it automatically transitioned to horizontalflight and landed back at the landing site where theballoon was released 145 minutes earlier. We will presentand analyzed flight test result from every phase of themission and summarize the lessons learned.

I. INTRODUCTION

Unmanned flying in the higher atmosphere hascome into focus in recent years. After decades ofmanned high altitude flying [X15, U2, SR71], keytechnology evolved in the 1990’s which mainlylead to the NASA family of high altitude solarplanes for example Pathfinder and Helios. As aresult of ongoing work, a flight of two weeks couldbe achieved by the Qinetic Zephyr airplane in2010, which is the reference to this date. Flying athigh altitude is necessary for these missions to stayabove the clouds. Flying below the clouds wouldprevent the plane from collecting sun energy. Sincealtitudes of 12 to 17km are not usable because ofhigh speed jetstream winds, usually an altitude of20km is targeted for flying. Even though this leadsto large, fragile airframe designs. For aerodynamicreasons and other environmental constraints, thealternative of low flying solar planes is only fea-sible for periods of low or no cloud coverage likedemonstrated by AtlanicSolar in 2015 [1].

M. Laiacker, S. Wlach and M. Schwarzbach are with the Roboticsand Mechatronics Center, German Aerospace Center (DLR),82230 Wessling, Germany, {maximilian.laiacker,sven.wlach, marc.schwarzbach}@dlr.de.

In recent years companies like Google and Face-book have started their own projects aiming forHALE solar planes supplying network servicesworldwide.

The technologies needed for the kind of HALE(high altitude long endurance) flights mentionedare manifold. All systems have to be extremelylightweight while still working under the low pres-sure and varying temperature. Operational aspectsof high altitude BLOS flying of UAV add to thechallenges. While many problems can be solvedanalytically or by testing in environmental simu-lation facilities, the final validation can only beachieved in the real environment.

Since HALE solar platforms are not (yet) avail-able and the construction is a major effort, theresearch group on flying robots of the GermanAerospace Center (DLR) chose an alternative ap-proach. Since high altitude balloon research is awell-established business and facilities are alsooffering segregated airspace up to the desiredaltitudes, the airplane for short duration testingcould be designed small and without the needof propulsion when lifted to 20km by a balloon.While it is not possible to test the effects of longduration exposure to altitude, results on aerody-namics, control, communication and operation canbe achieved by a small team using a smaller systemon a tight budget.The NASA ARES program [2]used a similar technique to simulate a plane flyingin Mars atmosphere.

II. UAV SYSTEM OVERVIEW

The UAV, shown in Figure 1, was specificallydesigned for this high altitude mission and hasa wingspan of 3m and a mass of 7.4kg. It isequipped with our modular autopilot system [3]but some new modules had to be developed for thismission. The long range manual control system

Fig. 1. The HABLEG UAV designed for high altitude missions

is also installed that allows manual flight usingremote control but not in a redundant configura-tion. Different to the modular concept describedin [3] to reduce weight and size of the overallsystem most of the electronic components havebeen integrated in one enclosure. When flying athigh altitudes, thermal design of electronic com-ponents is also very important because of the verycold environment temperatures and low densityatmosphere. More details of the thermal design andtesting of the autopilot system are presented in [4].

The central component of the autopilot is a Pow-erPC based flight control computer (FCC) that runsthe real time operating system QNX. The sensorsand actuators are connected via serial RS232 inter-face to the FCC. A small calibrated MEMS inertialmeasurement unit is used for attitude and headingestimation. Pressure and airspeed are provided byan airdata sonde. Two u-blox GPS receivers areinstalled. One is in the front and one is behind thewing inside the fuselage. Two spatially separatedand differently orientated U-Blox 6 GPS receiversare used for localization. They are voted for signalquality (number of satellites), which can greatlydiffer depending on orientation, due to the high lat-itude of the flight testing area. Beyond navigation,a valid GPS position is also critical to the flight

Fig. 2. The pan tilt unit with antennas used for the mission

termination logic, which in case of a telemetryand GPS failure would command the ejection of aparachute. The reception of GPS signals can easilybe jammed by internal or external RF sourcesso using two sensors spatially separated reducesthe risk that both sensors are affected. Only onesensor is used by the autopilot system and thedecision which one is done automatically basedon the reported number of received satellites andestimated position accuracy.

The airplane is also equipped with an analoguevideo downlink that is independent from the au-topilot system with the exception of the powersupply. This real time video downlink was added tobe able to easily and directly observe the aircraftsorientation and velocity during the mission by theoperator in the ground control station. Telemetrydata including sensor readings, estimated state andinternal states are transmitted over a low band-width bidirectional radio link to the ground stationand displayed to the operator. The operator canalso change parameters and issue commands to theUAV during the mission using this radio link.

The video downlink is using the S-Band at2.3GHz which is also used for telemetry onsounding rockets that are also launched at theESRANGE [5] test area where the mission wasplanned. Using this frequency allowed the usageof existing high gain tracking antenna and receiverinfrastructure at the site. For the telemetry data

radio link a dual frequency redundant system isused. The frequencies used are in the 2.4GHz and868MHz ISM band. Using the ISM bands insteadof restricted frequencies allows to use widelyavailable commercial components and to test themwithout special permits before the high altitudemission. On the ground a pan-tilt unit (PTU)is used to point high gain directional antennastowards the UAV during flight. The 21db parabolic2.4GHz and 13db yagi 868MHz antenna and PTUare shown in Figure 2. On the UAV side low gainomnidirectional antennas are used. The PTU iscontrolled based on the received UAV position andPTU position. During the mission there was alwaysa working communication link to the UAV.

The ground control station is set up to accom-modate three operators. The manual control oper-ator has a remote control transmitter and a screenshowing the analogue video image. The manualcontrol operator can switch to manual mode anytime if there is a working radio link and take overmanual control, for example in case the autopilotcontrol system fails. The tele command operatorcan send commands and parameter changes tothe autopilot control system. The tele commandoperator also observes the video downlink, artifi-cial horizon and position of the UAV on a map.The third operator is the flight test engineer andobserves all telemetry data and coordinates themission.

III. MISSION SIMULATION

To execute the mission automatically and fulfillall safety regulations, complex control softwarehad to be implemented and tested. To test thecontrol system a partial hardware in the loop setupwas implemented. In this setup the sensor data isgenerated by a software module that runs on theFCC instead of the sensor drivers. This is easilydone since a shared memory based middleware isused. During normal operation, several softwaredriver processes read and interpret the data comingfrom the connected sensors and provide these datausing the middleware to the other processes of theautopilot system. During a partial hardware in theloop setup the driver processes are stopped andsimulated driver data is provided by the simulationmodel. The commands for the actuators provided

by the control processes are read by the softwaredrivers that communicate with the real actuatorsand the simulation model. The advantages of thissimulation setup compared to pure software in theloop are that the motions of the actuators can beobserved and the telemetry and command interfaceto the system is the same than during a real flightexperiment. Since using this kind of simulation thehuman operator is part of the system the operatoris also tested and trained. This operator training isvery important for a successful mission since in acritical situation decisions and actions have to beexecuted fast. Mission checklists have also beenprepared and refined during these simulations.

The simulation model supports all phases of themission including the ascent when the UAV istethered to the balloon. Wind influences can alsobe simulated to create a realistic trajectory duringascent and the resulting headwind during the glidephase. The simulation is run at real time so thatthe operators know what to expect during the realmission. The simulation model also supports somefailure modes like a lost GPS signal. Communica-tion failures can be simulated by simply turning ofthe radio modem or removing the antennas. Themission phases can be summarized as follows:

• Ascent on balloon• Launch and drop• Transition to horizontal flight• High altitude flight experiments• Return to landing position• Landing

In Figure 3 data from a simulated transitionto horizontal flight is shown. After the horizontalflight is stabilized a course towards the landingposition is commanded. In this mission the landingposition is equal to the position where the UAVattached to the balloon was launched. During eachphase a different mode of the control systemis used and the transition between these phasesmust be detected automatically or activated by thesoftware. Low altitude flight test have also beenperformed. For these flight test the HABLEG UAVhas been equipped with an electric motor. Detailsabout the low altitude flight tests are described in[6].

0 10 20 30 40 50 60 70 80 90

−80

−60

−40

−20

0

20

40

60

80

Time[s]

hor. speed[m/s]vert. speed[m/s]Mode 1Mode 2Roll [Deg]Pitch [Deg]

Fig. 3. Simulated balloon launch phases, velocities and angles

aileron

1

roll angle limit

phi = atan (Vk/g*chi_dot )

chi_dot

Vkphiclc_Phi _des

K_chi

Integrator

1s

scaling

K_p

p

enable _roll _control

Ki_phi

K_phi

phi

enable _course_controlcourse ground _speed

desired course

1

Fig. 4. Block diagram of the course controller with mode selectionand gain scheduling implemented for the HABLEG mission

elevator

1

pitch angle limit

K_EAS

Integrator

1s

scaling

K_q

q

enable _pitch _control

Ki_theta

K_theta

theta

enable _EAS_controlEASdesired theta

2

desired EAS

1

Fig. 5. Block diagram of the pitch and EAS controller withmode selection and gain scheduling implemented for the HABLEGmission

IV. CONTROLLER DESIGN

The controller has to support all these differentflight phases in a wide range of true airspeed, alti-tudes and air densities. A cascaded PID controllerdesign is used. The inner loops control the rotationrates and angles. The elevator is used to control thepitch rate q and the pitch angle θ. The ailerons areused to control the roll rate p and the roll angleφ. The outer loops control the equivalent airspeedEAS and the course χ. In Figure 4 a block diagramof the horizontal controller for the course is shown.During the control mode "damp" the values for"enable course control" and "enable roll control"are zero. For the "roll zero" mode only "enableroll control" is set to one and for the "return tohome" and waypoint mode both are set to one. Theequivalent airspeed is controlled by changing thedesired θ angle. The airspeed loop can be disabledby the operator and an adjustable θ angle is used.This option was added because the performance ofthe used airspeed sensor was not known for highaltitudes. In Figure 5 the control block diagram forthe pitch angle and airspeed controller is shown.During the control mode "damp" the values for"enable EAS control" and "enable pitch control"are zero. During the transition to horizontal flightthe value for "enable pitch control" is slowlyincreased to one to make a slow transition froman almost vertical pitch angle to horizontal flight.When the airspeed control is manually enabledthe value for "enable EAS control" is set to one.The desired course or heading is calculated by thenavigation process or can also be entered manuallyby the operator. The desired θ angle is set by theoperator. To adapt the controller to the differenttrue airspeed and air density we added a gainscheduling based on the dynamic pressure for thecontrol surfaces which resulted in stability issueswhich will be discussed in the next section.

V. FLIGHT DATA ANALYSIS

In May 2014 a first campaign was conducted,which however didn’t provide a flight opportunitydue to unsuitable weather conditions. One yearlater, in May 2015, another - this time successful- attempt was made. The flight mission beganwith an overall uneventful but from the operatorpoint of view very tense balloon ascent where

Fig. 6. Ascent (red) and descent (green) flight path during theHABLEG mission

strong winds carried the UAV to a ground dis-tance of about 66km at the maximum altitudeof around 20km. Here the release command wasissued by the ground control station. During thedrop phase heavy oscillations occurred which hadto be addressed by adjusting control gains by tele-command. Since this was an anticipated behaviorand trained prior to the mission, the situation washandled by the ground crew within a few seconds,resulting in a safe transition to horizontal flight.After the high altitude test flight program hadended the glider entered an automatic landing pat-tern and finally landed close to the starting positionusing manual control. The mission trajectory isshown in Figure 6 and is summarized in Table I.

In the following the flight will be analyzed inmore detail.

A. Balloon ascent

HABLEG was attached to the balloon at endof the tail boom so it hang vertically, nose down.The balloon train ascended very stable and starteddrifting north due to southerly winds. During the

mission duration 145 minutesascent duration 75 minutes

max. altitude 19550mmin. altitude 330m

max. distance 66kmtotal distance traveled 169km

TABLE I

MISSION SUMMARY

0 5 10 15 20 25 300

2

4

6

8

10

12

14

16

18

20

[m/s]

Alti

tude

[km

]

wind velocityascent velocity

Fig. 7. Wind and ascent velocity of the balloon during ascent

ascent the autopilot was in "neutral" mode. Inthis mode the controller is basically switched offand all control surfaces stay in there neutral trimpositions.

As the balloon got higher the horizontal groundvelocity increased as the balloon entered thestrongest winds around 9km altitude. The windvelocity and ascent velocity profile is shown inFigure 7. An slight increase of ascent velocity withaltitude can also be observed. When an altitudeof 19.5km was reached the release command wasgiven.

B. Transition to horizontal flight

This is the most critical part of the mission.HABLEG has to transition from a vertical nose

Fig. 8. Still frame from the backward facing camera shortly afterreleasing HABLEG from the balloon

down attitude with almost zero airspeed to hori-zontal flight with around 50m

strue airspeed (TAS)

in very thin atmosphere. During the simulationthe transition happens like shown in Figure 3where the release command was given at timezero. With 90◦ pitch down it will pick up verticalspeed at 9.81m

s2and the controller in "neutral"

mode. A drogue chute will deploy to stabilize thedrop. After a fixed amount of 3s the vertical andhorizontal controller are switched to "damp" modeto dampen rotational velocities. When reaching avelocity of 35m

sthe drogue is released and the

horizontal controller controls the roll angle to zeroin "roll zero" mode and the vertical controller willslowly change the pitch angle to the desired valueof around −5◦. At 33s after the release commandthe start sequence is finished and the horizontalcontroller is switch to "return to home" which willinitiate a turn towards the landing position and thevertical controller is in "desired theta" mode.

During the real high altitude mission the transi-tion to horizontal flight needed small adjustmentsby the operators. The release from the balloonworked perfectly. The drogue chute deployed aswe could see afterwards on the video from thebackward facing video camera. A still frame ofthis moment is shown in Figure 8. Different tothe simulations the airplane started to oscillateafter the "damp" mode was activated. The controlgains used to dampen the rotations where too highso it had the opposite effect and lead to heavypitch and roll oscillations which can be seen in 9after 5s. These oscillations increased after the con-trollers are switched into the next modes "roll zero"

0 200 400 600 800 1000 1200 1400 1600 1800

−10

0

10

20

30

40

50

60

70

80

Time since release [s]

EAS[m/s]desired EAS[m/s]TAS[m/s]enable_speed_controlθ[deg]desired θ[deg]

Fig. 10. Airspeed and pitch controller performance during the firsthalf hour of the flight

and "desired theta" respectively. The operators inthe ground station recognized the oscillations andstarted to reduce the control gain Kp at 15s (Fig.11) and switched to manual control at 24s (Fig.9). When switched to manual control the controlswhere not used so the control surfaces stayed neu-tral and this immediately stopped the oscillationsand rotations of the UAV as shown in Figure 9after 25s. At 26s the autopilot is switched on again.The horizontal controller is now controlling the rollangle to zero as planned but the vertical controlleris still oscillating around the desired θ angle. Theautopilot is switched off a second time at 31s for 2swhile the control gain Kq for the vertical controlleris also reduced. At 33s the autopilot is switched onand the controller initiates a stable turn towards thelanding position. The turn is finished at 55s afterthe release command was given and HABLEG isthen flying stable at around 18km altitude towardsthe landing position still over 60km away.

C. High altitude flight

After the turn towards the landing position wasfinished the airspeed started to drop. This wascaused by the selected desired θ angle which wasto shallow to maintain the needed airspeed. Thisdrop in airspeed started pitch angle oscillations.The oscillation was recognized by the operatorsand the autopilot was switched off again for ashort period of time to stop the oscillation and to

0 10 20 30 40 50 60 70

−150

−100

−50

0

50

100

150

Time since release [s]

release commandautopilot onφ[deg]θ[deg]vorizontal velocity[m/s]vertical velocity[m/s]

Fig. 9. Flight phases, velocities and angles shortly after release from the balloon

change gains and increase the desired θ angle ataround 80s as shown in Figure 10. After these ad-justments airspeed settled at around 17m

sEAS and

the autopilot was only switched off again shortlybefore the manual landing. After these controlissues where solved two manual course changeswhere commanded to evaluate the maneuverabilityduring high altitude flying. The course changes canbe seen on the map in shown in Figure 6. Thesecourse changes with a maximum bank angle of 21◦

were very smooth and worked as expected. About800s into the flight the operators verified that theairspeed sensor works reliably and the airspeedcontrol loop was activated. The airspeed controllerstabilized the airspeed very close to the set-point asshown in Figure 10 beginning 800s. The airspeedcontrol is implemented as a P controller that addsan offset to the manually commanded desired θangle so it cannot fully eliminate a steady stateerror. During the whole flight the operators triedto manually optimize the glide ratio by adjustingthe θ pitch angle or desired airspeed as can beseen in Figure 10 at 400s and 1500s. The optimalglide ratio set point changes due to constantlychanging air density and headwinds during thedescent towards the landing position.

During the flight and after analyzing therecorded data a slow roll angle oscillation wasdetected. These oscillations originate from a quan-tization problem of the aileron control servos. Theaileron control surfaces are quite big to providegood maneuverability at low altitudes but at higheraltitudes where a higher TAS is required to gener-ate enough lift the aileron deflections required arevery small. At high speeds the aileron trim positionfor level flight is between two quantized servopositions which generates a slow roll oscillationbecause the servo will only switch position whena sufficient roll angle error is reached. Smallercontrol surfaces would be better suited for a highaltitude UAVs with the downside of losing somemaneuverability at low altitudes. The landing posi-tion was reached at about 4km altitude and abovethe clouds. HABLEG entered an automatic landingpattern over the landing area. The final approachand landing was done under manual control tosafely avoid several obstacles in the landing area.HABLEG came to a stop just 80m away fromwhere it was launched 145minutes earlier.

D. Controller gain scheduling for high altitudeflights

To adapt the controller for high altitude flight again scheduling of the gain Kp and Kq was imple-mented based on the ratio of the reference dynamicpressure and the current dynamic pressure.

qref = TASref2ρref

2(1)

q (TAS, ρ) = TAS2ρ

2(2)

scaleq (TAS, ρ) =qref

q (TAS, ρ)(3)

with TASref = 15ms

and ρref = 1.225. Thisvelocity scaling was used because the torque pro-duced by a control surface deflection scales withthe dynamic pressure but for a low inertia planelike a small UAV the aerodynamic damping hasa big effect and results in a fast settling finalrotation velocity proportional to a control surfacedeflection and total air speed (TAS) [7]. So scalingthe control gains by dynamic pressure results inthe same gains when flying at the same EAS athigh altitudes but since the TAS is much higher theresulting rotation speeds are much higher, leadingto an unstable controller. In Figure 11 the effectivenormalized control gains are shown using dynamicpressure scaling and TAS scaling with

scaleTAS (TAS) =TASref

max (TASref , TAS)(4)

where "param. 6" and "param.11" are the gainvalues without scaling. The "param." values arechanged by the operator during the flight. At theend of the flight in lower altitudes the gain parame-ter are increased for better controller performance.For pressure scaling the parameters values thatwhere changed by the operator during the flightare used to calculate the normalized gains andfor TAS scaling the initial parameters are usedfor comparison. So using the TAS scaling wouldprobably result in a stable controller during thecritical release phase since the values are smallerthan the ones used after they were changed by theoperator.

0 10 20 30 40 50 60 70 80 90 100 110 120 130 140 1500

0.5

1

1.5

Time since release [s]

scale

TAS

Kp

param. 6K

q

param. 11

Fig. 11. Gain scheduling comparison based in TAS and dynamicpressure during the transition to horizontal flight

E. GPS receiver performance

Two u-blox GPS receivers model u-blox6SuperSense R© have been installed. Both have beenconfigured to the "4g airborne" mode. Both re-ceivers maintained a position fix during the holefight. In Figure 12 the measured altitudes fromboth GPS receivers and the calculated pressure al-titude and the number of satellites used for positioncalculation is shown. During the ascent when theUAVs is vertical the fluctuation in used satellitesis higher because of the rotation of the yaw angle.The GPS receivers are mounted so that they pointupward in normal flight so, while hanging on theballoon, they point towards the horizon. Duringthe slow rotation during the balloon ascent thereceivers have better reception from different partsof the horizon thus a different amount of satelliteswill be received. After the transition to stablehorizontal flight the number of used satellites ismore stable since now the orientation of the GPSantennas is more stable and the satellites receptionis only effected by the slow motion of the satellitesover the sky.

During the unstable phase of the transition tohorizontal flight shortly after the release the umberof used satellites fluctuate fast and also drop byhalf to only 5 but this is still enough to maintain a

−1000 −900 −800 −700 −600 −500 −400 −300 −200 −100 0 100 200 300 400 5005

7

9

11

13

15

17

19

20

Time since release [s]

pressure altitude[km]GPS1 altitude[km]GPS2 altitude[km]GPS1 sat. usedGPS2 sat. used

Fig. 12. GPS and pressure altitude readings during ascent anddecent between 15 and 20km

3D position fix. During the transition from unstabledrop to horizontal flight the UAV exceeded the 4glimit set in the velocity model of the receiversand the reported position now differs by a largemount in the altitude. For flight control during thehole flight the GPS2 is used. By comparing theshape of the pressure altitude and GPS altitudeduring this phase around 50s after release it can beconcluded that GPS2 has calculated the correct al-titude while GPS1 produced erratic measurementsand the receiver voting algorithm did a good job inselecting the right receiver during the flight. Themeasurements of GPS1 stabilized shortly after thisphase of high vertical acceleration.

The difference between the pressure altitude andGPS altitude gets bigger with altitude but thereis almost no difference between the two GPSreceivers. The drift of the pressure altitude canbe explained by small measurement errors by thepressure sensor which result in a big altitude errorin this low pressure area of the atmosphere. Be-tween −100s and −10s the data was not recordedinside the autopilot for unknown reasons.

VI. CONCLUSIONS

With the success of the HABLEG mission wehave shown that high altitude flights are possiblewith this small platform. It was shown that havingthe partial software in the loop simulation helps alot in creating reliable flight control software andin effective training for the operators to quickly

react to problems during the mission. A more accu-rate aerodynamic model would have been helpfulsince it probably would have helped to discoverthe wrong approach to the control gain schedulingused for the mission. The simulation model usedbehaved well with the gains that result in unsta-ble flight during the real mission. Having a realtime video available also proofed to be invaluablefor situational awareness. In combination with aneasy to use ground station interface and intensivetraining, this allowed for a quick response to theencountered problem. Future work will focus onintegrating the lessons learned from this missionfor further high altitude missions and a possiblefuture project for a solar powered high altitudelong endurance UAV.

VII. ACKNOWLEDGMENT

The authors thank SSC and DLR Moraba forsupporting this project.

REFERENCES

[1] P. Oettershagen, A. Melzer, T. Mantel, K. Rudin, R. Lotz,D. Siebenmann, S. Leutenegger, K. Alexis, and R. Siegwart, “Asolar-powered hand-launchable uav for low-altitude multi-daycontinuous flight,” in Robotics and Automation (ICRA), 2015IEEE International Conference on. IEEE, 2015, pp. 3986–3993.

[2] J.-C. Lede, R. Parks, and M. Croom, “High altitude drop testingin mars relevant conditions for the ares mars scout mission,”in 2nd AIAA "Unmanned Unlimited" Conf. and Workshop,September 2003.

[3] M. Laiacker, A. Klöckner, K. Kondak, M. Schwarzbach,G. Looye, D. Sommer, and I. Kossyk, “Modular scalablesystem for operation and testing of uavs,” in American ControlConference, 2013.

[4] S. Wlach, M. Shwarzbach, and M. Laiacker, “Dlr hableg -high altitude balloon launched experimental glider,” in 22ndESA PAC Symposium 2015, 2015.

[5] “Esrange Space Center SSC,” http://www.sscspace.com/esrange-space-center-3.

[6] M. Schwarzbach, S. Wlach, and M. Laiacker, “Modifyinga scientific flight control system for balloon launched uavmissions,” in Aerospace Conference, 2015 IEEE. IEEE, 2015,pp. 1–10.

[7] H. H. Hurt, Aerodynamics for naval aviators. Office of theChief of Naval Operations, Aviation Training Division, USNavy, 1960.