際際滷

際際滷Share a Scribd company logo

Communication and Distributed Control in Multi-Agent Systems - Preliminary Model of Micro-unmanned Aerial Vehicles (MAV) Swarms

Download as PPT, PDF
Fabio Ruini
Fabio Ruini

Presentation held during the fourth euCognition six-monthly meeting (Venice, 10-11 January 2008). For more details about this work, please have a look to: http://www.tech.plym.ac.uk/soc/research/ABC/plymav/

1 of 11
Downloaded 40 times
Communication and Distributed Control in Multi-Agent Systems Preliminary Model of Micro-unmanned Aerial Vehicle (MAV) Swarms Fabio Ruini Adaptive Behaviour and Cognition Research Group (ABC) School of Computing, Communications and Electronics University of Plymouth, UK VENICE (ITALY), 10-11 JANUARY 2008 4TH EUCOGNITION SIX-MONTHLY MEETING Acknowledgments: euCognition Network Action NA097-3 EOARD European Office of Aerospace Research and Development (Grant 073075)
Introduction This work focuses on the use of Multi-Agent Systems for the modelling of Micro-unmanned Aerial Vehicles (MAVs) in a distributed control task. The task regards a search scenario in the context of security and urban counter-terrorism. The main goals of this research are: to demonstrate how a MAVs swarm can successfully be made autonomous, using evolutionary and neural computation techniques; to study how the communication between the team-mates can impact the performance of these flying robots.
Simulated environment 610 m (710 pixels) 650 m (760 pixels) Obstacles distribution: (Picture found on: http://www.aquiva.co.uk/canarywharf ) Obstacle Width Height Starting X Starting Y Final X Final Y 0 130 82 180 44 310 126 1 130 82 180 150 310 232 2 95 190 408 38 503 228 3 55 230 643 17 698 247 4 110 90 280 213 390 303 5 90 95 231 263 321 358 6 83 93 417 266 500 359 7 133 213 577 247 710 460 8 50 95 109 403 159 498 9 90 140 231 400 321 540 10 90 142 419 392 509 534 11 45 88 74 498 119 586 12 49 38 274 564 323 602 13 100 55 244 602 344 657 14 35 18 430 534 465 552 15 90 123 420 552 510 675 16 33 59 387 620 420 679 17 53 171 592 482 645 653 18 44 65 213 698 257 763
AeroVironment WASP Block III *Source: AeroVironment web site http://www.avinc.com/downloads/WASP-III_datasheet_6_5_07.pdf Technical specifications*: Length 38 cm (15 in) Wingspan 72 cm (28.5 in) Weight 430 g (0.95 lb) Speed 46-65 km/h (25-40 mph) Endurance 45 min Propulsion electric motor
Neural Network controller TARGET DISTANCE TARGET ANGLE ULTRA-SONIC PERCEPTION STEERING DETONATION Distance from the target Since the simulated environment can be seen as a Cartesian plane, the distance between the current MAV and the target is calculated as the Euclidean distance between their centre points and then discretized. Angle from the target The angle between the current MAV and the target is measured basing on its facing direction and then discretized using a Gray code-based coding. Ultra-sonic perception Each MAV is endowed with three ultra-sonic sensors, respectively oriented at 0属, 45属 and 315属 (-45属), able to detect the presence of an obstacle until 30 pixels far. Steering The output neuron dedicated to the steering is a continuous neuron whose output value ranges from -1 (340属/-20属) to +1 (+20属). Detonation The output neuron dedicated to the detonation is a Boolean neuron. 0 : do nothing; 1: detonate.
Simulator: a quick overview Neural network computational capability: neurons distribution along the various layers Genetic algorithm: number of seeds, number of generations (for a single seed), number of swarms belonging to the population, number of tests to be carried out on every swarm Genetic operators: mutation probability, mutation amount Various controls: enable/disable graphics view, background and statistics visualizations, manual control of the visualization speed, pause, zoom level, visualization speed Control buttons: start/stop evolution, start/stop test, exit
Some swarms in action End of the evolution Lets see how the MAVs perform: Beginning of the evolution
Evolution resume
Conclusions and future works In this preliminary model we have demonstrated how a neural network controller for MAVs swarms can be successfully evolved through multi-agent systems. Plans for future work: Use movable target, robust target; Study different communication capability/protocols (including self-emergent lexicons); Develop 3D simulator (physics library).
Contacts, links and publications Contacts: e-mail: [email_address] web: http://www.fabioruini.eu Skype: fabio.ruini Links: Project home page: http://www.tech.plym.ac.uk/soc/research/ABC/plymav / Adaptive Behaviour and Cognition Research Group: http://www.tech.plym.ac.uk/soc/research/ABC/ Publications: Whitepaper: http://www.eucognition.org/network_actions/NA097-3_outcome.pdf
End...

More Related Content

Communication and Distributed Control in Multi-Agent Systems - Preliminary Model of Micro-unmanned Aerial Vehicles (MAV) Swarms

  • 1. Communication and Distributed Control in Multi-Agent Systems Preliminary Model of Micro-unmanned Aerial Vehicle (MAV) Swarms Fabio Ruini Adaptive Behaviour and Cognition Research Group (ABC) School of Computing, Communications and Electronics University of Plymouth, UK VENICE (ITALY), 10-11 JANUARY 2008 4TH EUCOGNITION SIX-MONTHLY MEETING Acknowledgments: euCognition Network Action NA097-3 EOARD European Office of Aerospace Research and Development (Grant 073075)
  • 2. Introduction This work focuses on the use of Multi-Agent Systems for the modelling of Micro-unmanned Aerial Vehicles (MAVs) in a distributed control task. The task regards a search scenario in the context of security and urban counter-terrorism. The main goals of this research are: to demonstrate how a MAVs swarm can successfully be made autonomous, using evolutionary and neural computation techniques; to study how the communication between the team-mates can impact the performance of these flying robots.
  • 3. Simulated environment 610 m (710 pixels) 650 m (760 pixels) Obstacles distribution: (Picture found on: http://www.aquiva.co.uk/canarywharf ) Obstacle Width Height Starting X Starting Y Final X Final Y 0 130 82 180 44 310 126 1 130 82 180 150 310 232 2 95 190 408 38 503 228 3 55 230 643 17 698 247 4 110 90 280 213 390 303 5 90 95 231 263 321 358 6 83 93 417 266 500 359 7 133 213 577 247 710 460 8 50 95 109 403 159 498 9 90 140 231 400 321 540 10 90 142 419 392 509 534 11 45 88 74 498 119 586 12 49 38 274 564 323 602 13 100 55 244 602 344 657 14 35 18 430 534 465 552 15 90 123 420 552 510 675 16 33 59 387 620 420 679 17 53 171 592 482 645 653 18 44 65 213 698 257 763
  • 4. AeroVironment WASP Block III *Source: AeroVironment web site http://www.avinc.com/downloads/WASP-III_datasheet_6_5_07.pdf Technical specifications*: Length 38 cm (15 in) Wingspan 72 cm (28.5 in) Weight 430 g (0.95 lb) Speed 46-65 km/h (25-40 mph) Endurance 45 min Propulsion electric motor
  • 5. Neural Network controller TARGET DISTANCE TARGET ANGLE ULTRA-SONIC PERCEPTION STEERING DETONATION Distance from the target Since the simulated environment can be seen as a Cartesian plane, the distance between the current MAV and the target is calculated as the Euclidean distance between their centre points and then discretized. Angle from the target The angle between the current MAV and the target is measured basing on its facing direction and then discretized using a Gray code-based coding. Ultra-sonic perception Each MAV is endowed with three ultra-sonic sensors, respectively oriented at 0属, 45属 and 315属 (-45属), able to detect the presence of an obstacle until 30 pixels far. Steering The output neuron dedicated to the steering is a continuous neuron whose output value ranges from -1 (340属/-20属) to +1 (+20属). Detonation The output neuron dedicated to the detonation is a Boolean neuron. 0 : do nothing; 1: detonate.
  • 6. Simulator: a quick overview Neural network computational capability: neurons distribution along the various layers Genetic algorithm: number of seeds, number of generations (for a single seed), number of swarms belonging to the population, number of tests to be carried out on every swarm Genetic operators: mutation probability, mutation amount Various controls: enable/disable graphics view, background and statistics visualizations, manual control of the visualization speed, pause, zoom level, visualization speed Control buttons: start/stop evolution, start/stop test, exit
  • 7. Some swarms in action End of the evolution Lets see how the MAVs perform: Beginning of the evolution
  • 9. Conclusions and future works In this preliminary model we have demonstrated how a neural network controller for MAVs swarms can be successfully evolved through multi-agent systems. Plans for future work: Use movable target, robust target; Study different communication capability/protocols (including self-emergent lexicons); Develop 3D simulator (physics library).
  • 10. Contacts, links and publications Contacts: e-mail: [email_address] web: http://www.fabioruini.eu Skype: fabio.ruini Links: Project home page: http://www.tech.plym.ac.uk/soc/research/ABC/plymav / Adaptive Behaviour and Cognition Research Group: http://www.tech.plym.ac.uk/soc/research/ABC/ Publications: Whitepaper: http://www.eucognition.org/network_actions/NA097-3_outcome.pdf