CS580 Module Presentation

Transcript

1. CMPSC580 Module III Paper Summary Brandon Ginoza Andreas Landgrebe Cody Kinneer Jake Ballinger Hawk Weisman Department of Computer Science Allegheny College March 3, 2015
2. What are Agents? Definition (Agent) A computer system capable of autonomous action in an environment
3. What are Agents? Definition (Agent) A computer system capable of autonomous action in an environment 1. Autonomous: Designed to meet an objective
4. What are Agents? Definition (Agent) A computer system capable of autonomous action in an environment 1. Autonomous: Designed to meet an objective 2. Flexible, reactive, proactive, and social
5. What are Multi-Agent Systems? Definition (Multi-Agent System) A set of agents that interact to solve a problem in an environment 1. Communication 2. Interaction: cooperative or competitive
6. What are Distributed Systems? Definition (Distributed System) A collection of autonomous agents that communicate through a network 1. Coordinate activities 2. Share resources
7. Outline 1. The COMRADE System for Multi-Robot Autonomous Landmine Detection in Post-Conflict Regions 2. Cooperative perimeter surveillance with a team of mobile robots under communication constraints
8. COMRADES
9. COMRADES Cooperative
10. COMRADES Cooperative Multi-Robot
11. COMRADES Cooperative Multi-Robot Automated
12. COMRADES Cooperative Multi-Robot Automated Detection
13. COMRADES Cooperative Multi-Robot Automated Detection System
14. COMRADES: Goals Detect mines in post-conflict regions Low-cost Fault-tolerant distributed system
15. COMRADES: System Robots: 1. Explorer 2. Corobot
16. COMRADES: System Robots: 1. Explorer 2. Corobot Sensors: Metal Detector (MD) Infrared-Based Multi Detector (IR) Ground-Penetrating Radar (GPR)
17. COMRADES: Three Steps 1. Discover the free space in the environment. 2. For each robot, determine order to perform tasks. 3. Fuse the information at an object of interest obtained by different robots.
18. COMRADES: Algorithms Distributed area coverage Multi-robot task allocation (MRTA) Distributed information fusion
19. COMRADES: Algorithms Definition (Coverage Path Planning) “[A]llowing robots to plan their paths so that they can cover the entire free space of their environment using their coverage sensors.” Voronoi Partition-based Coverage
20. COMRADES: Algorithms Definition (Coverage Path Planning) “[A]llowing robots to plan their paths so that they can cover the entire free space of their environment using their coverage sensors.” Voronoi Partition-based Coverage Divide environment by Voronoi partition for each robot
21. COMRADES: Algorithms Definition (Coverage Path Planning) “[A]llowing robots to plan their paths so that they can cover the entire free space of their environment using their coverage sensors.” Voronoi Partition-based Coverage Divide environment by Voronoi partition for each robot Robots divide partitions into cells, cover with Spanning Tree Coverage
22. COMRADES: Algorithms Definition (Coverage Path Planning) “[A]llowing robots to plan their paths so that they can cover the entire free space of their environment using their coverage sensors.” Voronoi Partition-based Coverage Divide environment by Voronoi partition for each robot Robots divide partitions into cells, cover with Spanning Tree Coverage Obstacle avoidance / repartitioning
23. COMRADES: Algorithms Definition (Coverage Path Planning) “[A]llowing robots to plan their paths so that they can cover the entire free space of their environment using their coverage sensors.” Voronoi Partition-based Coverage Divide environment by Voronoi partition for each robot Robots divide partitions into cells, cover with Spanning Tree Coverage Obstacle avoidance / repartitioning Fault tolerance
24. COMRADES: Algorithms Definition (Multi-robot task allocation) “[H]ow to find a suitable assignment of tasks to robots so that the tasks performed by the robots can be completed in an efficient manner in terms of time and energy expended by the robots.”
25. COMRADES: Algorithms Definition (Multi-robot task allocation) “[H]ow to find a suitable assignment of tasks to robots so that the tasks performed by the robots can be completed in an efficient manner in terms of time and energy expended by the robots.” Each type of sensor must visit a potential landmine
26. COMRADES: Algorithms Definition (Multi-robot task allocation) “[H]ow to find a suitable assignment of tasks to robots so that the tasks performed by the robots can be completed in an efficient manner in terms of time and energy expended by the robots.” Each type of sensor must visit a potential landmine Equivalent to multiple traveling salesman problem (NP-Hard)
27. COMRADES: Algorithms Definition (Multi-robot task allocation) “[H]ow to find a suitable assignment of tasks to robots so that the tasks performed by the robots can be completed in an efficient manner in terms of time and energy expended by the robots.” Each type of sensor must visit a potential landmine Equivalent to multiple traveling salesman problem (NP-Hard) Heuristic solution using a spatial task queue
28. COMRADES: Algorithms: Multi-robot task allocation 1. Intertask transition matrix M(t) =     π11 π12 . . . π1n π21 π22 . . . π2n . . . πn1 πn2 . . . πnn    
29. COMRADES: Algorithms: Multi-robot task allocation 1. Intertask transition matrix M(t) =     π11 π12 . . . π1n π21 π22 . . . π2n . . . πn1 πn2 . . . πnn     2. Robot state vector Vri (t) =      πi1(t)) πi2(t)) . . . πin(t))     
30. COMRADES: Algorithms: Multi-robot task allocation 1. Intertask transition matrix M(t) =     π11 π12 . . . π1n π21 π22 . . . π2n . . . πn1 πn2 . . . πnn     2. Robot state vector Vri (t) =      πi1(t)) πi2(t)) . . . πin(t))      3. Task proximity vector ´ Vri (t) = Vri (t) × M(t)
31. COMRADES: Algorithms: Multi-robot task allocation 1. Intertask transition matrix M(t) =     π11 π12 . . . π1n π21 π22 . . . π2n . . . πn1 πn2 . . . πnn     2. Robot state vector Vri (t) =      πi1(t)) πi2(t)) . . . πin(t))      3. Task proximity vector ´ Vri (t) = Vri (t) × M(t) 4. Robot spatial task queue
32. COMRADES: Algorithms: Multi-robot task allocation Robots announce a bid for the task at the top of the queue Robot with highest bid (lowest distance) is assigned task Robots announce when they have completed a task, transition matrices rebuilt
33. COMRADES: Algorithms Definition (Information aggregation) “[T]o combine the information about the characteristics of a potential landmine from different types of sensors and make a decision whether the object is indeed a landmine”
34. COMRADES: Algorithms Definition (Information aggregation) “[T]o combine the information about the characteristics of a potential landmine from different types of sensors and make a decision whether the object is indeed a landmine” Performed by prediction market technique
35. COMRADES: Algorithms Definition (Information aggregation) “[T]o combine the information about the characteristics of a potential landmine from different types of sensors and make a decision whether the object is indeed a landmine” Performed by prediction market technique Sensor agent calculates probability that an object is a landmine
36. COMRADES: Algorithms Definition (Information aggregation) “[T]o combine the information about the characteristics of a potential landmine from different types of sensors and make a decision whether the object is indeed a landmine” Performed by prediction market technique Sensor agent calculates probability that an object is a landmine Sensor agent calculates report based on belief and expected reward
37. COMRADES: Algorithms Definition (Information aggregation) “[T]o combine the information about the characteristics of a potential landmine from different types of sensors and make a decision whether the object is indeed a landmine” Performed by prediction market technique Sensor agent calculates probability that an object is a landmine Sensor agent calculates report based on belief and expected reward Sensor agent submits report to market maker agent
38. COMRADES: Algorithms Definition (Information aggregation) “[T]o combine the information about the characteristics of a potential landmine from different types of sensors and make a decision whether the object is indeed a landmine” Performed by prediction market technique Sensor agent calculates probability that an object is a landmine Sensor agent calculates report based on belief and expected reward Sensor agent submits report to market maker agent Market maker agent calculates aggregate belief from reports
39. COMRADES: Algorithms Definition (Information aggregation) “[T]o combine the information about the characteristics of a potential landmine from different types of sensors and make a decision whether the object is indeed a landmine” Performed by prediction market technique Sensor agent calculates probability that an object is a landmine Sensor agent calculates report based on belief and expected reward Sensor agent submits report to market maker agent Market maker agent calculates aggregate belief from reports Decision maker agent decides what other sensors should be deployed
40. COMRADES: Future Work Sensor devices & environmental conditions
41. COMRADES: Future Work Sensor devices & environmental conditions Improve coverage algorithm
42. COMRADES: Future Work Sensor devices & environmental conditions Improve coverage algorithm Using aerial robots
43. COMRADES: Future Work Sensor devices & environmental conditions Improve coverage algorithm Using aerial robots
44. COMRADES: Improving Coverage Algorithm Sharing maps between robots
45. COMRADES: Improving Coverage Algorithm Sharing maps between robots Planning coverage maps
46. COMRADES: Improving Coverage Algorithm Sharing maps between robots Planning coverage maps Improved information fusion techniques
47. COMRADES: Improving Coverage Algorithm Sharing maps between robots Planning coverage maps Improved information fusion techniques
48. Sharing maps between robots Multi-Robot Coverage with Dynamic Converage Information Compression Done: Reduced algorithm overhead
49. Sharing maps between robots Multi-Robot Coverage with Dynamic Converage Information Compression Done: Reduced algorithm overhead To Do:
50. Sharing maps between robots Multi-Robot Coverage with Dynamic Converage Information Compression Done: Reduced algorithm overhead To Do: Better than randomly selecting a new area after failing to bound a polyregion
51. Sharing maps between robots Multi-Robot Coverage with Dynamic Converage Information Compression Done: Reduced algorithm overhead To Do: Better than randomly selecting a new area after failing to bound a polyregion Dynamically adjust the the block size according to robot’s coverage performance
52. Sharing maps between robots Multi-Robot Coverage with Dynamic Converage Information Compression Done: Reduced algorithm overhead To Do: Better than randomly selecting a new area after failing to bound a polyregion Dynamically adjust the the block size according to robot’s coverage performance Combine perceived long-range sensor data with historical coverage information
53. Sharing maps between robots Multi-Robot Coverage with Dynamic Converage Information Compression Done: Reduced algorithm overhead To Do: Better than randomly selecting a new area after failing to bound a polyregion Dynamically adjust the the block size according to robot’s coverage performance Combine perceived long-range sensor data with historical coverage information Lossless compression techniques for coverage data to increase the accuracy of fused covered regions
54. COMRADES: Future Work Sensor devices & environmental conditions
55. COMRADES: Future Work Sensor devices & environmental conditions Improve coverage algorithm
56. COMRADES: Future Work Sensor devices & environmental conditions Improve coverage algorithm Using aerial robots
57. COMRADES: Future Work Sensor devices & environmental conditions Improve coverage algorithm Using aerial robots
58. COMRADES: Future Work Relative Localization Approach for Combined Aerial and Ground Robotic System Aerial robots help ground robots more effectively navigate the environment
59. COMRADES: Future Work Relative Localization Approach for Combined Aerial and Ground Robotic System Aerial robots help ground robots more effectively navigate the environment Done: Widens the area range for ground robots
60. COMRADES: Future Work Relative Localization Approach for Combined Aerial and Ground Robotic System Aerial robots help ground robots more effectively navigate the environment Done: Widens the area range for ground robots To Do: Obstacles
61. COMRADES: Future Work RQ: In order to best share information (and therefore avoid obstacles on the ground), should aerial robots act as the information aggregator for a team of ground robots where it concerns map sharing?
62. Cooperative perimeter surveillance with a team of mobile robots under communication constraints
63. Cooperative perimeter surveillance with a team of mobile robots under communication constraints 1. A perimeter P := {p(x) ∈ R2 : x ∈ [0, L]}
64. Cooperative perimeter surveillance with a team of mobile robots under communication constraints 1. A perimeter P := {p(x) ∈ R2 : x ∈ [0, L]} 2. A set of N mobile robots Mi ∈ {M1...Mn}
65. Cooperative perimeter surveillance with a team of mobile robots under communication constraints 1. A perimeter P := {p(x) ∈ R2 : x ∈ [0, L]} 2. A set of N mobile robots Mi ∈ {M1...Mn} 3. Each robot can move along P in two directions with a variable speed vi
66. Cooperative perimeter surveillance with a team of mobile robots under communication constraints 1. A perimeter P := {p(x) ∈ R2 : x ∈ [0, L]} 2. A set of N mobile robots Mi ∈ {M1...Mn} 3. Each robot can move along P in two directions with a variable speed vi 4. Communication radius of robots is not wide enough to cover the entire perimeter
67. Perimeter Surveillance: Strategies Frequentist strategies
68. Perimeter Surveillance: Strategies Frequentist strategies Goal: Optimize for refresh time, uniformity, or average frequency
69. Perimeter Surveillance: Strategies Frequentist strategies Goal: Optimize for refresh time, uniformity, or average frequency Cyclic: robots spaced evenly across perimeter
70. Perimeter Surveillance: Strategies Frequentist strategies Goal: Optimize for refresh time, uniformity, or average frequency Cyclic: robots spaced evenly across perimeter Non-frequentist strategies
71. Perimeter Surveillance: Strategies Frequentist strategies Goal: Optimize for refresh time, uniformity, or average frequency Cyclic: robots spaced evenly across perimeter Non-frequentist strategies Stochastic or non-deterministic
72. Perimeter Surveillance: Strategies Frequentist strategies Goal: Optimize for refresh time, uniformity, or average frequency Cyclic: robots spaced evenly across perimeter Non-frequentist strategies Stochastic or non-deterministic Behavioral control with finite state automata
73. Perimeter Surveillance: Strategies Frequentist strategies Goal: Optimize for refresh time, uniformity, or average frequency Cyclic: robots spaced evenly across perimeter Non-frequentist strategies Stochastic or non-deterministic Behavioral control with finite state automata Many previous strategies are centralized
74. New Strategy Frequentist: optimizes for refresh time
75. New Strategy Frequentist: optimizes for refresh time Decentralized: one-to-one exchange of coordination variables
76. New Strategy Frequentist: optimizes for refresh time Decentralized: one-to-one exchange of coordination variables Path partitioning: each robot covers a non-overlapping segment Partitions vary in length based on robot speed
77. Perimeter Surveillance: Algorithm All robots move at max speed Meet neighbors at segment endpoints — continuing past endpoint if neighbor not found Share coordination variables: sum of speeds and perimeter length
78. Perimeter Surveillance: Algorithm Robust against failure At most 2N communications to converge to optimum solution Algorithm verified by simulation and a team of robots
79. Perimeter Surveillance: Simulations Simulations: conducted in MATLAB
80. Perimeter Surveillance: Simulations Simulations: conducted in MATLAB Proposed solution compared with other one-to-one coordination techniques
81. Perimeter Surveillance: Simulations Simulations: conducted in MATLAB Proposed solution compared with other one-to-one coordination techniques Varied: Perimeter length Number of robots
82. Perimeter Surveillance: Simulations Simulations: conducted in MATLAB Proposed solution compared with other one-to-one coordination techniques Varied: Perimeter length Number of robots Maximum speeds distributed uniformly
83. Perimeter Surveillance: Experiments Experiments: conducted with Pioneer 3-AT Robots
84. Perimeter Surveillance: Experiments Experiments: conducted with Pioneer 3-AT Robots Convergence time and robustness: tested by causing robots to join and leave the team
85. Perimeter Surveillance: Experiments Experiments: conducted with Pioneer 3-AT Robots Convergence time and robustness: tested by causing robots to join and leave the team Adaptability: size of perimeter varied during experiment
86. Perimeter Surveillance: Results Simulation:
87. Perimeter Surveillance: Results Simulation: Both algorithms converged at solution
88. Perimeter Surveillance: Results Simulation: Both algorithms converged at solution Convergence time of the proposed approach scales better with number of robots
89. Perimeter Surveillance: Results Simulation: Both algorithms converged at solution Convergence time of the proposed approach scales better with number of robots Experiments:
90. Perimeter Surveillance: Results Simulation: Both algorithms converged at solution Convergence time of the proposed approach scales better with number of robots Experiments: Performance conformed to predictions
91. Perimeter Surveillance: Future Work Heterogenous terrain and locomotion Varied sensor types Perimeters with priority Staggered outages
92. Conclusion COMRADES Cooperative Multi-Robot Automated Detection System Goals: Detect mines in post-conflict regions, low cost, fault-tolerant The Steps: Discover free space, determine order, fuse information Algorithms: Coverage path planning, multi-robot task allocation, and information aggregation Future Work: Combining aerial and ground robots
93. Conclusion Perimeter Surveillance New strategy: optimizes for refresh time, decentralized, and better path partitioning Algorithm: Robust, at most 2N communications, and experimentally verified Future Work: Heterogenous terrain, priority, or varied sensor types?
94. Bibliography [1] Jose Joaquin Acevedo, Bego˜ na C Arrue, Ivan Maza, and An´ ıbal Ollero. Cooperative perimeter surveillance with a team of mobile robots under communication constraints. In Intelligent Robots and Systems (IROS), 2013 IEEE/RSJ International Conference on, pages 5067–5072. IEEE, 2013. [2] Thumeera R Wanasinghe, George KI Mann, and Raymond G Gosine. Relative localization approach for combined aerial and ground robotic system. Journal of Intelligent & Robotic Systems, 77(1):113–133, 2015. [3] Zachary Wilson, Taylor Whipple, and Prithviraj Dasgupta. Multi-robot Coverage with Dynamic Coverage Information Compression. University of Nebraska at Omaha, 2014.