Design of Autonomous Navigation Controllers for Unmanned Aerial Vehicles using Multi-objective Genetic Programming Gregory J. Barlow March 19, 2004
Overview Background Unmanned Aerial Vehicle Control Evolution and Fitness Evaluation Experiments and Results Conclusions and Future Work
Evolutionary Computation Biologically inspired computational method of problem solving May be applied to a variety of structures (binary strings, real numbers, computer programs, hardware, neural networks, etc) because the algorithm operates on an encoding of the parameters, not the parameters themselves
Genetic Programming A population of random programs is created Each individual in the population undergoes a fitness test and is assigned a fitness value Genetic operators (crossover, mutation, etc) are performed on the population to form the next generation The process is repeated until a suitable individual is evolved
Evolutionary Process Population Selection Evaluation Parent(s) Genetic Operator Child(ren)
Representation Each individual is a program, which we represent as a tree Function set: for non-leaf nodes Terminal set: for leaf nodes
Crossover
Mutation
Unmanned Aerial Vehicle Control Create controllers that will fly a UAV toward a target radar and then circle the radar for jamming Make the UAV controller completely autonomous Be able to handle multiple radar types Be able to transfer evolve controllers to real UAVs
Simulation To test the fitness of a controller, the UAV is simulated for 4 hours of flight time in a 100 by 100 square nmi area The initial starting position of the UAV is randomly set along the bottom of the simulation space The position of the radar is also randomly set for each simulation UAVs can sense the AoA and amplitude of incoming radar signals
Simulation
Transference These controllers should be transferable to real UAVS. To encourage this: Only the sidelobes of the radar were modeled Noise is added to the modeled radar emissions The angle of arrival value from the sensor is only accurate within ±10°
Functions and Terminals Hard Left, Hard Right, Shallow Left, Shallow Right, Wings Level, No Change IfThen, IfThenElse, And, Or, Not,, >=, > 0, < 0, =, +, -, *, / Amplitude > 0, Amplitude Slope 0, AoA
Fitness Functions Normalized distance Circling distance Level time Turn cost
Normalized Distance
Circling Distance
Level Time
Turn Cost
Performance of Evolution Multi-objective genetic programming produces a Pareto-optimal front of solutions, not a single best solution. To gauge the performance of evolution, fitness values for each fitness measure were selected for a minimally successful controller.
Baseline Values Normalized Distance0.15 Circling Distance4 Level Time1000 Turn Cost0.05
Direct Evolution Experiments Continuously emitting, stationary radar Intermittently emitting, stationary radar with a regular period Intermittently emitting, stationary radar with an irregular period Continuously emitting, mobile radar Intermittently emitting, mobile radar with a regular period
Direct Evolution Radar Type RunsControllers TotalSucc.RateTotalAvg.Max. Continuous, Stationary %3, Intermittent, stationary (regular period) %1, Intermittent, stationary (irregular period) %2, Continuous, mobile %2, Intermittent, mobile (regular period) %
Continuously emitting, stationary radar
Circling Behavior
Intermittently emitting, stationary (regular)
Intermittently emitting, stationary (irregular)
Continuously emitting, mobile radar
Intermittently emitting, mobile radar
Incremental Evolution Continuously emitting, stationary radar (seed populations) Intermittently emitting, stationary radar Continuously emitting, mobile radar Intermittently emitting, stationary radar (multiple increments) Intermittently emitting, mobile radar (multiple increments)
Incremental Evolution Radar Type RunsControllers TotalSucc.RateTotalAvg.Max. Continuous, Stationary %2, Intermittent, stationary %2, Continuous, mobile %2, Intermittent, stationary (multiple increments) %2, Intermittent, mobile (multiple increments) %1,
Intermittent, mobile (multiple increments)
Transference to a wheeled mobile robot Controllers were designed for UAVs A UAV was not yet available for flight tests to evaluate transference Evolved controllers were tested on a wheeled mobile robot, the EvBot II A speaker was used in place of the radar, and an acoustic array in place of the radar sensor
EvBot II PC/104 processor Communications with a wireless network card Runs Linux On-board acoustic array
Transference considerations In simulation, the sensor accuracy was ±10°, but the acoustic array accuracy was approximately ±45° Wheeled robot not controlled by roll angle, must be turned and then moved The size of the maze environment was not equivalent to the simulation environment, instead the scale size of the maze environment was 1.13 by 0.9 nautical miles
Sensor accuracy Sensor accuracy of ± 10°Sensor accuracy of ± 45°
Controller 1
Controller 2
Conclusions Autonomous navigation controllers were evolved to fly to a radar and then circle around it while maintaining stable and efficient flight dynamics Multi-objective genetic programming was used to evolve controllers Controllers were evolved for five radar types using both direct evolution and incremental evolution
Conclusions Incremental evolution dramatically increased the success rates for the more difficult radar types Methods were used to aid in transference of controllers to real UAVs Controllers were tested on a wheeled mobile robot with good success Evolved controllers are capable of transference to real physical vehicles
Future Work Controllers will be tested on physical UAVs for several radar types Distributed multi-agent controllers will be evolved to handle cases of multiple UAVs against multiple radars Incremental evolution will be used to aid in the evolution of fit multi-agent controllers for complex radar types