denardi sab06 srw

50 %
50 %
Information about denardi sab06 srw

Published on January 13, 2008

Author: Ubert


Slide2:  Key ideas The helicopter platform Evolutionary control system design Ongoing work: state estimation Future work: model identification Slide3:  A flock of instrumented helicopters, capable of performing cluster computation and coordinated flight. Flight coordination based on flocking rules. Cluster computation exploiting wide band, short range wireless communication. Slide4:  Main focus of the ongoing development: Helicopter platform design Helicopter control issues Flocking algorithm The field of mobile cluster computing is still characterized by open issues. In the meantime an approach based on distributed computation is proposed. Slide5:  Produce a group of indoor flying vehicles (with complex dynamics) that demonstrates fluid aerial flocking to: Explore the dependence between vehicle dynamics and flocking rules Investigate the role of the sensors Slide6:  Aerial mapping/imagery Surveillance Environmental sensing Just to name a few…. Slide7:  Cylindrical arena 10 m diameter, 6 m height Slide8:  Lighter than air (blimp): easy to control; low size/payload ratio; simple vehicle dynamics; long endurance. Small aircraft (slow flyer): relatively easy to control; low size/payload ratio; minimum forward speed; good endurance. Rotorcraft (helicopter, quadrirotor): not easy to control; complex dynamics; good size/payload ratio; sufficient flight endurance. Slide9:  Hirobo Lama XRB model: compact and efficient coaxial design flybar controlled rotor to augment stability maximum flight time of about 15 minutes full three dimensional maneuverability maximum payload of about 40 - 50g Slide11:  Characteristics: 32bit ARM7TDMI 40MHz microcontroller 2 independent I2C ports (IMU, ultrasonic sensors) 2 serial ports (reflashing, Gumstix comm) 2 external interrupts (rotor speed encoders) same dimensions of the original electronics, but lighter 6 configurable leds Slide12:  Bluetooth / Wifi data link (fly by wire / data link with flockmates and base station) Single Board Linux Computer (400MHz Intel Xscale, 64MB ram, 4MB flash) Slide13:  6 DoF inertia measurement unit (accelerations, magnetic field, and angular speeds) 3 D infrared tracking system (indoor global position) Vertical ultrasonic range sensor (take off/ landing) Slide14:  Hirobo helicopter retrofitted with the newly designed electronics, the Gumstix SBC and the IMU. All the controls can be operated from a remote PC. Slide15:  Characteristics: frame rate 200 Hz 1 mm spatial accuracy hundreds of markers can be tracked simultaneously markers position, rigid bodies position and attitude streamed in real-time (delay < 10ms) Only limited sensor information will be relayed wirelessly e.g. : relative bearing and distance obstacles information limited FOV Slide16:  problem General automated method to design the low level controller (i.e. below the flocking layer) solution Neuroevolution not available yet! Dynamic model of the real helicopter replacement freely available qualitatively similar dynamic helicopter simulator state controls Slide17:  The dynamic simulator accepts the same control inputs of a real helicopter and provides the helicopter state (every 0.02s): velocity in body coordinates (u,v,vz) [and distance from waypoint (Dx,Dy,Dz)] angular attitude (f,q,y) and angular speeds (p,q,r) u p, q, r, v vz y f q Slide18:  Evolution Strategy ES (10+23), used to evolve the synaptic weights of a network with predefined topology. population of 33 networks rank based selection: the worst 23 individuals are replaced by mutated copies of individuals from the elite self-adaptive mutation was preferred to simple Gaussian mutation in some of the runs typically run for 500 generations no crossover Slide19:  A straightforward topology (i.e. MLP) is difficult to train using a GA. Very brittle solutions can be obtained with backpropagation. Neural interference and genetic interference are likely to be the cause of this. Solution Incremental evolution can provide some sort of “bootstrap” by evolving Yaw stabilization beforehand in a separate task A modular topology inspired by the dynamic symmetry of the helicopter model was used Slide20:  Reaching and keeping a randomly generated commanded velocity [-3.5 , 3.5] ft/s for a defined amount of time [12 , 350] timesteps . Fitness inversely proportional to the error between the commanded and the instantaneous helicopter velocity. u v vz Slide21:  Conservative overdamped speed control Steady state error Slide22:  Flying a path of randomly generated waypoints (in the 3D space), in a given amount of time... Fitness proportional to the number of waypoints traveled, penalty for not flying the shortest path and not keeping the reference heading. Slide23:  The best evolved controller performing the waypoint following task. Mean distance of 6.82 m between two consecutive waypoints. Each black dot marks the position of the helicopter every 10 timesteps (0.2 s). Slide24:  Estimation obtained with an UKF, integrating gyros measurements and using acceleration and magnetometer readings for correction. Attitude estimation is essential to stabilize the helicopter… Slide25:  Continuous time kinematic equations driven by IMU measurements angular velocities. Quaternions representation of attitude. Estimation of the gyro bias. Gyros integration… Slide26:  The triaxial magnetometer and the accelerometers readings provide the observation data. An additive noise scheme is assumed. The acceleration feedback is disengaged when the magnitude of the accelerations vector is different from |g|. Correction from accelerometers and magnetometers… Slide27:  In order to cope with the non linearity of the system model, an Unscented Kalman Filter is preferred to a classic EKF. The implementation is simplified since the derivatives of the state equations do not need to be computed; the square root formulation is effectively less computationally expensive than a standard EKF. The filter is expected to run on the embedded ARM7 processor with a reasonable update frequency (e.g. 50Hz) Slide28:  An accurate model of the helicopter dynamic is not available but… very accurate flight data is available from the 3D tracking system. Proposed approach Extract the dynamic model in the acceleration space by using a machine learning technique. Slide29:  Integrating the acceleration equations produces a non linear model. The model does not include any information about the platform, and is computationally not expensive. (source Abbeel et al. 2005) Slide30:  Renzo De Nardi, Julian Togelius, Owen E. Holland and Simon M. Lucas (2006). Evolution of Neural Networks for Helicopter Control: Why Modularity Matters. To be presented at the IEEE Congress on Evolutionary Computation. Renzo De Nardi, Owen Holland, John Woods, and Adrian Clark (2006). Swarmav: A swarm of miniature aerial vehicles. Proceedings of the 21st Bristol International UAV Systems Conference. O. Holland, J. Woods, R. De Nardi, A. Clark. Beyond swarm intelligence: The Ultraswarm. Proceedings of the IEEE Swarm Intelligence Symposium (SIS2005), eds. P. Arabshahi and A. Martinoli, IEEE, June 2005

Add a comment

Related presentations