330 likes | 339 Views
On swarm robotics. A beginner ’ s view. Luboš Popelínský Knowledge Discovery Lab Faculty of Informatics, Masaryk University Brno popel@fi.muni.cz, http://www.fi.muni.cz/kd. Overview. 1. Introduction to swarm intelligence 2. Swarm robots: Perception and communication
E N D
On swarm robotics. A beginner’s view Luboš Popelínský Knowledge Discovery Lab Faculty of Informatics, Masaryk University Brno popel@fi.muni.cz, http://www.fi.muni.cz/kd
Overview 1. Introduction to swarm intelligence 2. Swarm robots: Perception and communication Swarm robotics control algorithms 3. Temporal logic: formal specification of emergent behaviours in swarm robotics systems. Temporal and spatiotemporal refinement operator Appendix 1: Learning when to auction and when to bid
Swarm intelligence based on the collective behavior of decentralized, self-organized systems population of simple agents interacting locally with one another and with their environment follow very simple rules It leads to the emergence of complex global behavior Bee hive, ant colonies, bird flocking, animal herding, bacterial growth, and fish schooling Bonabeau E. Thrauluz G. Dorigo M. Swarm Intelligence Oxford University Press 1999 Introduced by Gerardo Beni and Jing Wang in 1989, in the context of cellular robotic systems.
Bees Beehive metaphor Foraging, randomly at the begining or Learning in the hive: dancing floor and auditory Web search Schultze, S.J.: A Collaborative Foraging Approach to Web Browsing Enrichment. In:Proc. CHI 2002, ACM, 2002, 860-861. Lorenzi F. Sherer dos Santos D. Bazzan A.L.C. Negotiation for task allocation among agents in case-based recommender systems. In IJCAI-05 Ws on Multiagent IR and Recommender Systems
Ant colony and source allocation Without a control center Without direct communication between ants An ant is building a path If succeeds to find a source, follows the same path back and sign it with a pheromone. The shorter path, the higher level of pheromone = positive feedback Consequence: more and more ants follow the most promising pathes
Ants colony and source allocation How it correspond to classification? source = learning examples from the same class path = between nodes <attribute=value> result = classification rule A1=v1 ^ A2=v2 ^ … ^ An=vn => class
Swarm robotics multirobot system which consist of large numbers of simple physical robots A key-component = communication between the members of the group that build a system of constant feedback local communication - wireless, e.g. bluetooth or infrared
Related fields Multi-agent systems Swarm intelligence Robotics Sensor networks But new…
Why do we need microrobots? can provide accurate handling of micro and nano parts exempt humans from tedious and very lengthy tasks can be used in hazardous environments can be cheaper to build than equipment currently used provide flexible and programmable systems for microassembly 'encourage' the development of novel manipulating tools
Swarm robotics algorithms Dispersion in indoor environments Distributed localization and mapping Mobile formation Cooperative hole avoidance Don Miner, Swarm Robotics Algorithms (2007)
Dispersion in indoor environments Uniform dispersion Wall nodes, frontier nodes (both do not move), interior nodes Disperse robots uniformly Generate vectors away from N particular neighbors Explore boundaries Frontier node send a message so that each node knows a number of hops from a frontier Then an interior node moves towards the lowest numbered neighbor (fastest path to the frontier)
Distributed localization and mapping Main idea: robot-beacons - are broadcasting position information Move in general direction IF num. of beacons goes below a threshold THEN become beacon IF num. of dependent nodes goes below a threshold THEN stop being a beacon, return to (1.)
Mobile formation Moving a large number of robots while maintaining connectivity Model: newtonian physics Force=mass*acceleration and Lennard-Jones (LJ) forces (modelling crystalline formation, liquids, gases) Results: with LJ performed much better
Cooperative hole avoidance Clearance sensors - to detect holes Traction sensors - to detect movements of other S-bots Evolutionary algorithm used Drawbacks: evolutionary algorithms are very slow -learing is done in simulation, not work in real environment: robots would fall in holes
Summary Local communication A robot usually represented by a finite state automaton Easy to represent in first-order logic: A robot + a context (neighbors) Probabilitic automata, e.g. Markov chains ? Temporal, spatiotemporal logic?
Temporal logic and swarm robotics Temporal logic for formal specification (and proving) emergent behaviour of a robotic swarm Allan Winfield,…, Michael Fisher. On Formal Specification of Emergent Behaviours in Swarm Robotic Systems, Intl. Journal on Advanced Robotic Systems Vol 2., p. 363-371
Basic algorithm Range-limited wireless communication rw - radius for communication ra - collision avoidance radius - number of neighbors threshold Default = forward moving, transmitting ‘I am here’ If num.of neighbors < (moving out of the swarm) Then turn 180 If num.of neighbors > (regained) Then execute a random turn
A linear time temporal logic Discrete time, linear ordering s0, s1, s2, s3, s4, … Finite past, infinite future Modalities NEXT <formula> SOMETIMES <formula> ALLWAYS <formula>
Simplified algorithm Robots move in a grid world to N(orth), E(ast), S(outh), W(est) Can turn before making a move 90 right, 90 left, 180 back Robots always move units Avoidance state is omitted = 1
State transition Forward state, connected -> move forward Forward state, not connected -> turn 180, change state to Coherent Coherent state, not connected -> move forward Coherent state, connected -> perform random turn (90 right, 90 left), change state to Forward
Temporal logic and beyond Specification expressed in First-order temporal logic (FOTL) mapping to monodic FOTL (max. 1 free variable) TeMP - resolution-based theorem prover for FOTL Refinement operator for temporal logic exists even for spatiotemporal logic (Blaťák, Popelínský, ECML’04 WS)
Learning when to auction and when to bid Market based approach frequently used for multi-robot coordination task=good, robots bid in auction for these goods Communication cost = number of messages needed for running the auctions Computational cost = cost of running the auctions Here: learning to reduce communication and computation cost Learning the probability of whether a given bit may win an auction D. Busquets, R. Simmons (CMU) DARS’06
Learning when to auction and when to bid II Usually bidders respond to all the tasks being announced Here: Compute Prob, probability of a bid being awarded in an auction Generate R, a random number. IF IR < Prob THEN bid Similar for the tasks auctioned Off-line learning
Experiments To characterize a set of rocks at different locations 3 settings NP (no probability), AuP (auction), AllP (auction and bid) AuP: Num. of rocks same, much better performance: auctions: 1394 -> 350 AllP: Num. of rocks slightly smaller, much lower cost messages: 13606(8608) -> 3814 Challenge = on-line learning