EXPERIMENTAL DEMONSTRATION OFCOORDINATED CONTROL FOR

MULTI-VEHICLE TEAMS

Ellis King, 1 Mehdi Alighanbari, 2 Jonathan How 3

Aerospace Controls LaboratoryMassachusetts Institute of Technology

Abstract: This paper introduces two unique testbeds that have recently beendeveloped at MIT to demonstrate coordination and control algorithms for teamsof autonomous UAVs. The first testbed uses eight rovers and four blimps operatedindoors to emulate a heterogeneous fleet of vehicles that could be used toperform a search and rescue mission. The second testbed uses eight small aircraftthat are flown autonomously using a commercially available autopilot. Thiscombination of testbeds provides platforms for both advanced research and veryrealistic demonstrations. Numerous trajectory optimization and team coordinationalgorithms have recently been developed to perform these missions. This paperhighlights several algorithmic pieces of the system and typical results are presentedfor representative experiments in each of the testbeds. Copyright c© 2004 IFAC

Keywords: Task Assignment, Receding Horizon Control, Multi-vehicle Teams

1. INTRODUCTION

UAVs offer advantages over conventional mannedvehicles in many applications because they canbe used in situations otherwise too dangerousfor manned vehicles (e.g., eliminating anti-aircraftdefenses) and without being weighed down bythe systems required by a pilot, UAVs are ca-pable of staying aloft on longer surveillance mis-sions. While the roles and capabilities of UAVsare growing, current UAV control structures wereconceived with limited roles in mind for thesevehicles. Thus it is necessary to improve on thiscontrol structure in order to fully exploit theexpanding capabilities of UAVs. This includesdeveloping techniques to optimize the coordina-tion of a fleet of UAVs, which is comprised ofthe coupled subproblems of determining sub-teamcomposition, allocating resources, and optimizing

1 Research Assistant, etking@mit.edu2 Research Assistant, mehdi a@mit.edu3 Associate Professor, MIT 33-328, 77 Mass. Ave.,

Cambridge, MA 02139. jhow@mit.edu

vehicle trajectories (Chandler, 2002). These are allcomputationally intensive optimization problemsthat require good situational awareness to achievecoordinated and cooperative behaviors. Numer-ous algorithms have recently been developed toachieve this cooperative behavior (Alighanbari etal., 2003; Richards et al., 2003), but a key step to-wards transitioning these high-level algorithms tofuture missions is to successfully demonstrate thatthey can handle the implementation challengesusing scaled vehicles operating in realistic environ-ments. Thus two testbeds have been developed atMIT to perform these demonstrations. One usesmultiple rovers and blimps operated indoors toemulate a heterogeneous fleet of vehicles perform-ing a search and rescue mission. The second useseight UAVs that are flown autonomously usinga commercially available autopilot from CloudCap Technology. Performing experiments on thesetestbeds will highlight the fundamental challengesassociated with: (i) planning for a large teamin real-time; (ii) developing controllers that arerobust to uncertainty in situational awareness,

Actuator input

Assign-ments

Minimum turn radiusNominal speed

UAVmodel

Vehicle statesObstaclesTargets

Vehicle statesObstacles

Waypoints and Activities

Vehicle capabilityNominal speed

Inertialsensordata

Sensor Measurements

ObstaclesTargets

Approx. Cost

Low levelController

TrajectoryDesigner

Graph- basedPath planning

Task Assignment

communications

Vehicle states

World Estimates

UAV Planner

Inter-vehicle

Predictor / Comparator

Vehicle / Simulation

VehicleStates

Uncertainty( )

Fig. 1. System algorithm architecture for rover, blimp and UAV testbeds.

and are sufficiently flexible to respond to dynamicchanges; and (iii) using communication networksand distributed processing to develop integratedand cooperative plans.

2. COORDINATION ALGORITHMS

Mixed-integer Linear Programming (MILP) haspreviously been shown to provide a natural frame-work for posing coordination problems, and ap-proximations such as the decomposition approachhave proven to provide accurate yet tractablesolutions to the overall problem (Bellingham etal., 2001; Richards et al., 2002). The decompo-sition approach simplifies the coupling betweenthe assignment and trajectory design problemsby calculating and communicating only the keyinformation that connects them. This is achievedusing an approximate cost-to-go calculation toobtain good estimates of the costs associated withfeasible paths around “obstacles” (e.g. buildings,no-fly-zones) in the environment. These costs arethen used in the assignment problem solved usingthe petal algorithm (Laporte and Semet, 2002;Richards et al., 2002). Uncertainty in the targetclassification due to poor or conflicting informa-tion enters the problem as uncertainty in the as-signment costs. As demonstrated by (Bertuccelliet al., 2004), the assignment process must be ro-bust to these types of uncertainty, and it is alsovital to ensure that the reconnaissance and striketasks are allocated simultaneously to provide themost benefit to the strike part of the missions.

Task Assignment - While the decomposition ap-proach greatly reduces complexity, the problemof task assignment with precedence constraintshas been shown to be NP-Hard, and achievingthe exact solution for a large team of UAVs iscomputationally intensive and not suitable forreal-time applications. The petal algorithm uses aheuristic method to prune out the solutions thatare not likely to be part of the optimal solution,

which significantly speeds up the task assignmentprocess. While this approach has been shown towork well for small problems, it is still difficult torun in real-time for problems with larger numbersof vehicles and tasks. To perform reassignmentin real-time as required in a dynamic, frequentlychanging environment, we have extended the petalmethod to develop a receding horizon task assign-ment (RHTA) algorithm.

RHTA significantly reduces the computation timeby selecting at most m (typically less than 3) tasksfor each UAV, then repeating the optimizationover several iterations. The process selects tasksto the mission list for each UAV, updates theUAV’s position and time, removes the assignedtasks from the task list, and repeats until allthe tasks are assigned. Timing and precedenceconstraints can be imposed using the approachin (Richards et al., 2002) or by simply requiringthat tasks be removed from the list until allprecedents have been assigned. The first of thesetwo approaches is more complicated, but shouldyield a less conservative approach than the second.

Trajectory Optimization - The final step is tocompute detailed UAV trajectories around theobstacles, which can be solved using a MILP-based receding horizon planner (Bellingham etal., 2002). This approach has been proven toguarantee the arrival at the target in boundedtime, as RH-MILP uses a simple vehicle dynamicsmodel in the near term and an approximate pathin the long term. This combination gives a goodestimate of the cost-to-go and greatly reducesthe computational effort required to design thecomplete trajectory. Discrepancies in the assump-tions made in the two models are handled byensuring that the planning horizon is sufficientlylong (Kuwata and How, 2004). Novel pruning andgraph search algorithms have recently been inte-grated with RH-MILP, and these also have theeffect of significantly reducing the computationalload. In the following experimental sections, the

Fig. 2. Control architecture: Decentralized pathplanning/Distributed task assignment.

Fig. 3. Control architecture: Decentralized pathplanning/Centralized task assignment.

trajectories shown are designed in real-time usingthis approach.

Control Architecture Figure 1 shows the controlarchitecture used for the UAVs, but the setup isvery similar for the rovers and blimps (Richards etal., 2002; Richards et al., 2003). Low-level controland the basic estimation tasks are run onboard,and the planning for the vehicles is done off-board. The planner outputs dynamically feasiblewaypoint lists and actions (i.e. classify, strike,assess) to the vehicles, and monitors the uncertainstates of the vehicles and the world map. Whensignificant changes to the situational awarenessare detected, the cost map is then updated, thetasks are re-assigned and/or the trajectories areredesigned.

Note that the system infrastructure was set upto emulate a fully integrated fleet of UAVs – alldata passes through a central hub that performsdata management between the planning comput-ers and vehicles, effectively simulating communi-cation delays, vehicle sensors and uncertainty inthe environment. Using this setup greatly simpli-fies the testbed, while maintaining nearly all ofthe functionality of a fully integrated system. Forexample, as shown in Figures 2 and 3, we can useour testbed to investigate the impact of communi-cation networking issues on the coordination prob-lem by imposing various limitations/constraintson how the planning laptops communicate (usingtheir own wireless or Ethernet links). Future workwill demonstrate the effectiveness of various con-trol architectures on the task assignment process,as would be seen in utilizing dynamic sub-teamsof various compositions.

The following sections describe the two multi-vehicle testbeds that have been developed to in-

vestigate the performance of these coordinationand control algorithms.

3. HARDWARE TESTBEDS

Rover/Blimp Testbed – The first testbed usesmultiple rovers and blimps operated indoors toemulate a heterogeneous fleet of vehicles thatcould be used to perform Suppression of EnemyAir Defense (SEAD) type missions. The roversin Figure 4 are ActivMedia’s P3-ATs, which areoperated with minimum speed and turn rate con-straints to emulate the motion of an aircraft.A Sony VAIO mounted on the rover processessensor data and performs the low-level control,while all high-level planning is done off-boardusing 2.4 GHz Dell laptops running MATLAB,AMPL, and CPLEX. A direct wireless Ethernetconnection provides a fast and reliable networkbetween the laptops, so this is essentially the sameas having both laptops onboard. The ArcSecondConstellation 3D-i is used to measure the vehicleposition indoors. This sensor uses laser metrologyto provide ±4mm position accuracy at 20Hz. The7ft diameter blimps in Figure 5 were scaled tocarry the VAIO and have an identical controlarchitecture. The blimps were designed to performreconnaissance and classification tasks in conjunc-tion with the rovers that act as strike vehicles.The blimps can also be used to map an uncertainenvironment for the rovers.

Figures 6-8 show the result for a scenario with 4rovers and 14 tasks, which introduces enough com-plexity in the mission that a global optimizationwould be prohibitive for dynamic real-time reas-signment. However, RHTA with m = 2 is shown toprovide good performance despite the complexityof the mission scenario. The mission begins with4 rovers and initial assignments, but after severalseconds, rover 4 is lost. Figure 6 shows the effect ofreassignment when the tasks of rover 4 have beendistributed among the other rovers. Later in themission, reassignment must again be performedwhen tasks 9 and 11 are determined to be locatedat different positions than previously assumed.Figure 7 shows the final assignments when twonew tasks (15 and 16) are discovered and rover 3is assigned to visit these locations.

Figure 9 shows an experimental result on therover/blimp heterogenous testbed. The scenariomodels the environment where the informationis partially available, requiring search missionsbe executed concurrently with the strike. Theobstacles impede the rovers, but the blimp is ableto fly over them. In this experiment, the blimp isused to search for tasks while the rover executesthem. The trajectory on the left shows the pathof the blimp performing its search pattern, and

Fig. 4. 4 of 8 ActivMedia P3-AT Rovers.

Fig. 5. 1 of 4 Blimps.

the trajectory on the right shows the path ofthe rover as it navigates to each task. For thepurpose of this test, the heading command issent to the blimp, while the rover follows thewaypoints (marked with ©) generated by theRHTA. Initially, the waypoint C was not known,but the blimp is sent to perform a reconnaissanceof the open space to the left, flying over theobstacle in the bottom of the figure. While therover is en route to execute task B, the blimpdiscovers the new task (waypoint C) and RHTAreassigns that task to the rover. This initial resultdemonstrates the successful integration of theheterogenous vehicles in our planning system andfuture tests will incorporate robustness into thetask assignment for a heterogeneous system ofrovers and blimps in an uncertain environment.

UAV Testbed – The second testbed is a fleet of 8UAVs (Figure 10) that are flown autonomouslyusing the Cloud Cap commercial autopilot in-terfaced directly with the planning and task as-signment algorithms. Figure 11 shows the 7.5ozPiccolo autopilot from Cloud Cap Technologiesinstalled in the aircraft. Small aircraft (60-sizedtrainers) were purposefully chosen to reduce op-erational complexity while still providing a highdegree of flexibility in the missions that can beperformed. The large trainer wing and Saito-91four-stroke engine allow an additional two poundsof payload for sensor or communications upgrades.Twenty minute flights are easily achievable in the

0 2 4 6 8 10 12 14 16 18 20

0

5

10

15

20

25

30

1

2

3

4

5

6

7

8

911

13

14

1 2 34

x[m]

y[m

]

Fig. 6. Plan after first re-assignment (rover 4 islost)

0 2 4 6 8 10 12 14 16 18 20

0

5

10

15

20

25

30

1

2

3

4

5

6

7

8

911

13

14

15

16

1

2

3

4

x[m]

y[m

]

Fig. 7. Last assignment (new tasks at wpts 15 and16 appear)

0 2 4 6 8 10 12 14 16 18 20

0

5

10

15

20

25

30

X [m]

Y [m

]

Fig. 8. Rover trajectories as measured during theexperiment

0 5 10 150

2

4

6

8

10

12

14

16

18

20

22

X [m]

Y [m

]

C

B

A

Blimp

Rover

UnknownWpt

Fig. 9. Rover and Blimp trajectories recorded dur-ing a heterogeneous recon and strike demon-stration.

current configuration, and further extensions arepossible.

The UAV testbed has been operated autonomouslyon numerous occasions – Figure 13 shows the re-sults of an 22 minute autonomous flight involvingtwo UAVs simultaneously flying the same flightplan. Both vehicles tracked the waypoints in the

Fig. 10. 6 of 8 UAVs.

Fig. 11. Piccolo autopilot from Cloud Cap Tech.

Fig. 12. Overhead of the local flying field at CrowIsland taken with the onboard camera.

presence of wind and open loop formation flightwas achieved by adjusting the commanded speeduntil the vehicles were in phase with one another.A 50 meter altitude offset was applied to one ofthe vehicle trajectories in Figure 13 to allow foreasier viewing.

A wireless video system has been integrated withthe UAV testbed to produce high quality imagesfrom the airborne vehicles – Figure 12 shows atypical aerial shot from one of the UAVs. Thissystem is used to verify the position of the vehiclesand provide user feedback for high level decisionmaking.

Figures 14–16 show results from a mission flownon the UAV testbed using receding horizon controlto generate waypoint plans in real time. In thisscenario, the goal locations used in the planner(shown as circles) were set in a 400m × 600m

−71.496−71.494

−71.492−71.49

−71.488−71.486

42.417

42.418

42.419

42.42

42.421

140160180200

Latitude

Longitude

Alt

[m]

Veh 1Veh 2 + 50m OffsetWaypoints

Fig. 13. Autonomous UAV flight data.

box pattern over the extents of the flying fieldand timing constraints (Alighanbari et al., 2003)were enforced to ensure a clockwise sequencing.Figures 14–16 show the progression of the optimalplanned paths and the telemetry data from thevehicle during one circuit. Each “×” segment rep-resents one waypoint plan which is returned fromthe optimal trajectory designer and is uploadedpiecewise to the UAV as the MILP optimizationcompletes. The plans are spaced at 100 metersor approximately 4 seconds from the next in thesequence. With 2.4 GHz Dell laptops runningCPLEXv8.0, these computations typically takeless than 1 second providing substantial marginin completion times.

The gray regions in Figures 14–16 represent obsta-cle locations which were encoded into the scenariousing mixed integer constraints (Bellingham etal., 2002) to constrain the planned trajectorieswithin a safe operating distance. Although nodynamic changes were made to the environment inthis scenario (i.e. pop-up obstacle, goal discovery),the framework of the RH controller also allows forthis capability.

This flight test was conducted in the presence ofwinds approximately 25% of the vehicle airspeed(The PT-60 aircraft are sized for 25 m/s nom-inal airspeed) and in a confined area, requiringminimum turn radius constraints to be enforced.As a result of these conditions, roughly 40 meterovershoot offsets can be seen in the flown trajec-tories in some instances. Wind disturbances actas a large contributor to the position offsets seenin Figure 16, as such current research investigatesthe incorporation of along and cross track errorfeedback into the receding horizon formulation forimproved performance. In addition, along trackspeed control is implemented to ensure meetinglocal timing constraints and coordination amongmembers of the fleet.

Planning Computers and Networks – There are12 Dell planning laptops connected on a 100 MbpsEthernet (the “wired” network). A wireless Ether-net access point is used to connect the Dells with

0 200 400 600 800 1000

−100

0

100

200

300

400

500

600

[m]

[m]

T = 41 sec

UAV 1RH TrajectoryCurrent GoalNext Goal

Fig. 14. Demonstration of UAV receding horizoncontrol 41 seconds into the circuit. Plannedpaths are the dashed lines, vehicle telemetryshown by solid lines.

0 200 400 600 800 1000

−100

0

100

200

300

400

500

600

[m]

[m]

T = 58 sec

Fig. 15. 58 seconds into the circuit.

0 200 400 600 800 1000

−100

0

100

200

300

400

500

600

[m]

[m]

T = 75 sec

Fig. 16. Completion of the circuit- Wind distur-bances are ∼ 5 m/s (W to E).

the VAIOs on the rovers (802.11b at 11 Mbps).All of the main pieces of software (controller,MATLAB) use TCP/IP to communicate, which iseasy to use and maintain. The Dell computers arealso equipped with a 10-seat license for CPLEXV8.0 which is used by the high-level planningalgorithms. Additionally, there is a second wire-less network at 54 Mbps based on the 802.11gstandard that connects the base station to fourvideo cameras on the rovers. There are also fourSRM6000 (Data-Linc Group) radio modems forlonger range communications (≈ 5km).

4. CONCLUSION

This paper presents hardware demonstrations ofthe receding horizon task assignment and trajec-tory design on two new rover/blimp and UAVtestbeds. These multi-vehicle testbeds provideunique platforms to evaluate various distributedcoordination and control strategies. Future workwill integrate distributed collision avoidance for-mulations, task assignment with the formation ofdynamic sub-teams, and missions with heteroge-neous vehicles (e.g. several rovers and blimps).

5. ACKNOWLEDGMENTS

Research was funded by AFOSR Grant #F49620-01-1-0453 and the testbeds were funded by DURIPGrant #F49620-02-1-0216.

REFERENCES

Alighanbari, M., Y. Kuwata, and J. How (2003).Coordination and control of multiple uavswith timing constraints and loitering. Amer-ican Control Conference.

Bellingham, J., A. Richards and J. How (2002).Receding horizon control of autonomousaerial vehicles. American Control Conferencep. 37413746.

Bellingham, J., M. Tillerson, A. Richards andJ. How (2001). Multi-task assignment andpath planning for cooperating uavs. Confer-ence on Cooperative Control and Optimiza-tion,.

Bertuccelli, L., M. Alighanbari and J. How (2004).Robust planning for coupled, cooperative uavmissions. submitted to the Conference on De-cision and Control.

Chandler, P. (2002). Complexity in uav coopera-tive control. American Control Conference.

Kuwata, Y. and J.P. How (2004). Stable trajec-tory design for highly constrained environ-ments using receding horizon control. IEEEAmerican Control Conference.

Laporte, G. and F. Semet (2002). Classical heuris-tics for the capacitated vrp. In: The VehicleRouting Problem (P. Toth and D. Vigo, Eds.).SIAM. Philadelphia.

Richards, A., J. Bellingham, M. Tillerson andJ. How (2002). Co-ordination and control ofmultiple uavs. AIAA Guidance, Navigation,and Control Conference.

Richards, A., Y. Kuwata and J.P. How (2003).Hardware demonstration of real-time milpcontrol. AIAA Guidance, Navigation, andControl Conference.