Distributed Intelligent Systems W12: Distributed Building and. and Artificial Systems

Transcription

1 Distributed Intelligent Systems W12: Distributed Building and Self-Assembling in Natural and Artificial Systems

2 Distributed building Outline Stigmergy, environmental templates Natural and artificial examples Slf Self-assembling General background Artificial examples (sub-mm and cm scales)

3 Distributed Building in Natural Systems

4 Example of Wasp Nest Guy Theraulaz

5 Example of Wasp Nest (sagittal cut) Guy Theraulaz

6 Further Examples of Wasp Nests Guy Theraulaz

7 Further Examples of Wasp Nests Pascal Goetgheluck

8 Not only Wasps: Termite Nest are also Impressive m high - thermoregulated Masson

9 Coordination Mechanisms for the Building Activity

10 Hyphotheses for Distributed ib t Building Mechanisms The plan Environmental template Stigmergy

11 The Plan Hypothesis? Scott Camazine

12 Hyphotheses for Distributed Building Mechanisms 1 The plan 2 3 Environmental template Stigmergy

13 The Environmental Template lt Hypothesis The building plan pre-existsexists in the environment under the form of spatial heterogeneities. The social insect activity only outlines these pre-existing environmental template. Environmental changes performed by insects play a minimal role in the building activity itself. This isa form of centralized if information seen from the bildi building insects t perspective There are several forms of environmental template: gradients naturally existing in the environment (humidity, temperature ) chemical gradients generated by one or more (static) individualsid of the colony

14 Ex.: Nest Structure in the Ant Acantholepis Custodiens Eggs T Larvae Pupae 3.00 a.m p.m.

15 Ex.: Termite Nest built around the Queen

16 Hyphotheses for Distributed Building Mechanisms 1 The plan 2 3 Environmental template - centralized Stigmergy

17 Stigmergy Definition It defines a class of mechanisms exploited by social insects to coordinate and control their activity via indirect interactions and stimulant signs left in the environment. Response R 1 R 2 R 3 R 4 R 5 Stimulus S 1 S 2 S 3 S 4 S 5 Stop time Stigmergic mechanisms can be classified in two different categories: quantitative (or continuous) stigmergy and qualitative (or discrete) stigmergy

18 Stigmergy The role of the two different stigmergic mechanisms in distributed building: 1 Sequence of stimuli and answers quantitatively different Quantitative (continuous) positive feedback and generation of spatial pattern mediated by self- organized agents 2 Sequence of stimuli and answers qualitatively different

19 Quantitative Stigmergy gy Features The successive stimuli quantitatively differentiate in their amplitude and merely modify the answer probability of other individuals Examples : mass recruitment in ants dead ant aggregation in ant cemetery pillar building in termites

20 Quantitative Stigmergy Pillar building in termites Spatial distribution of insects and their building activity are locally controlled by the pheromone o e density. Termites impregnate with pheromones the building material Pheromones diffuse in the environment Individuals carrying ground bullets follow the chemical gradient; they climb towards the highest pheromonal concentration and drop there their bullets Material drop rate is proportional to the number of active insects in the local region (positive feedback)

21 Simulation of Termite Nest Building: Pillar Bootstrapping

22 Stigmergy The role of the two different stigmergic mechanisms in distributed building 1 Sequence of stimuli and answers quantitatively different 2 Sequence of stimuli and answers qualitatively different Qualitative (discrete) positive feedback and assembly process mediated by self-organized agents

23 Qualitative Stigmergy gy Features Successive stimuli are qualitatively different. This process generates an agentmediated assembly dynamics (possibly including some self-alignment mechanisms of the bricks themselves) A B C S 1 S 2 S 3 time

24 Summary of Distributed Building Mechanisms in Natural Systems 1 The plan 2 3 Environmental template - centralized Stigmergy - distributed Blend!

25 Nest Building in Polist Wasps

26 Nest Building in Polist Wasps Guy Theraulaz

27 Nest Building in Polist twasps Organization of the building activity it It is indirectly carried out via the different, local configurations a wasp can find in the nest A probability of deposing a new cell is associated to each configuration

28 Nest Building in Polist twasps Potential sites for building S 2 S 1 S 2 S 1 S 1 S 2 S 1 S 3 S 1 S 1 S 1 S 2

29 Nest Building in Polist Wasps Probability of creating a new cell given the configuration of neighboring cells 1,0 0,8 0,6 0,4 0,2 0, Number of adjacent cell walls

30 Si l ti f th N t Simulation of the Nest Building Process

31 A Swarm on a 3D Lattice: The Nest Simulator lt A swarm of nest builders agent features Agents team is homogeneous No global plan of the whole building Local perception of the environment Random movements on a 3D lattice, no trajectories No embodiment (1agent = 1 single cell), no interference (however 2 agents cannot occupy the same cell) Reactive actions (applying (ppy rules) Distributed action during an iteration (n agents working together means n actions, not necessarily n dropped bricks); all n agents see the same overall nest configuration at a given iteration

32 The Nest Simulator Definition of agent s neighborhood for hexagonal cells The neighborhood of each agent is defined as the 20 cells around it (7 above, 6 around, and 7 below). z+1 z z - 1 z + 1 z z X A X Neighborhood Cell contents

33 The Nest Simulator Examples of building rules Z+1 z Building process is irreversible (no cell can be removed) Deposit rules can be strictly deterministic or probabilistic Z-1

34 Example of Rule Set Example of a rule set (Polist wasps, 7 rules)

35 Example of Nest Structure Obtained structures (Polists wasps, 7 rules) With deterministic rules With probabilistic rules

36 Looking for Stable Nest Architectures What kind of nest architectures can we build with the same method? What are the rules to implement in order to obtain stable architectures? An architecture is considered stable when several runs of a simulation with ihthe same rule set (but different random generator seed) generate architectures with the same global structure. Mechanisms of stigmergic coordination reduce the number of stable architectures.

37 Examples of Stable Architectures Parachatergus (21 rules) Vespa (13 rules) Stelopolybia (12 rules)

38 Examples of Stable Architectures Chatergus (39 rules) Artificial Nest Structure (35 rules)

39 Looking for Stable Nest Architectures How to build a stable architecture? The building process is carried out in successive steps. The current local configuration generate a stimulus different from that of the previous and of the successive configuration (qualitative stigmergy). Use of «bottleneck» rules for repeated pattern (ex. floor-to-floor coordination) Only this type of building algorithms generate coherent architectures. The whole set of these algorithms generates a limited number of nest shapes.

40 Reverse Engineering a Nest Structure using ECTechniques Fitness function: weighted sum of space filling and pattern replication arbitrary criteria based on 17 human observers who were asked to evaluate the amount of structure in a set of 29 different patterns, [Bonabeau 2000]) Biased evolution: Probabilistic templates (reduction of stimulating configurations containing a large number of bricks). Start microrule. No diagonal deposits (space not filled). Problems: Fitness function: Functional? Esthetical? Biological plausible? Mathematical definition? Disruptive effect of crossover; preservation of variable length blocks of rules (sequence of microrules needed for coordinated algorithms).

41 Distributed Building in Artificial Systems

42 2D Structure Building using Kheperas Robot behavior Reactive, non- communicating, nonadaptive behavior Quantitative stigmergy minimal (clear difference with Beckers et al 1994): the bigger, the more stable the cluster big cluster (> 2) = number of manipulation sites as cluster of 2 seeds almost no difference between cluster incrementing and decrementing probabilities Qualitative stigmergy gy important: 2 rules in interaction with cluster: Avoid if interaction with the cluster body Manipulate if interaction with cluster tips 1robot state: loaded, unloaded

43 Robot Controller Start Look for seeds Object detected? N N Y Carrying a seed? Y N Pick up the seed and decrement cluster s size Obstacle? Obstacle? Y Y N Obstacle avoidance Drop the seed and increment cluster s size

44 Parameter Characterization Start Look for seeds N Object detected? Probots +Pwalls +Pclusters (kt) Carrying a seed? 1 - [Probots +Pwalls + Pclusters (kt)] Y N Obstacle? Y Y Obstacle? Pick up the seed and decrement cluster s size Drop the seed and increment cluster s size Obstacle avoidance

45 Parameter Characterization Geometric Estimations Resulting Probabilities Incrementing probabilities Decrementing probabilities Perimeters are relevant for computing the cluster modifying probabilities: robot turns on the spot for object distinction before approaching the cluster!

46 Compare with Beckers et al. Geometric Estimations Resulting Probabilities Decrementing probabilities (by 1 puck, by 2 pucks) Decrementing probabilities (by 1 puck, by 2 pucks)

47 Robot Controller From Agassounon et al, 2004, correct representation:

48 Micro-Macroscopic Models From Agassounon et al, 2004, correct representation:

49 Models: Explanations and Predictions Single cluster? All models predicted yes and in roughly how much time! Number of clusters (inter-distance between seeds = 1seed) monotonically decreases if: Probability to create a NEW cluster of 1 seed in the middle of the arena is equal to zero No hard partitioning of the arena (robot homogeneously mix clusters) Cluster are not broken in two parts by removing one seed in the middle

50 Long Distributed Building Experiments Module-Based Model (Webots) Real robots (Khepera) 10% white noise on all sensor and actuators Perfectly homogeneous team Kinematic mode Electrical floor: continuous power supply in any position and orientation Heterogeneities among teammates and components Inaccuracies in acting and sensing Dynamics (e.g., friction) plays a role

51 Results (till single cluster) 3 robots real robots (5 runs), Webots (10 runs), AB-microscopic model (100 runs) [Martinoli, Ijspeert, Mondada, 1999] Mean size of clusters Size of the biggest cluster Number of clusters

52 Example of arising 2D Structures Noise in S&A and poor navigation capabilities do not allow for precise, controllable structure building Webots Real robots

53 Macroscopic Model: Distributed ib Building Dynamics d i (kt) = decr_geom_probability i *p_find i (kt) c i (kt) = incr_geom_probability i *pfind p_find i (kt) p_find i (kt) = finding probability of all the cluster of size i If n = number of seeds -> macroscopic model of environment with n nonlinearly coupled ODE (n for each possible cluster size) + robot states [Agassounon et al, Autonomous Robots, 2004]

54 Some Results from Agassonoun et al., 2004 (1 and 5 robots always active) Model comparison: macroscopic vs. submicroscopic models Mean Cluster Size Mean Number of Clusters

55 Some Results from Agassonoun et al., 2004 (10 robots always active) Model comparison: macroscopic vs. submicroscopic (Webots) Mean Cluster Size Mean Number of Clusters

56 Self-Assembling

57 What is Self-Assembly? Spontaneous organization of pre-existing units into spatial structures or patterns without external guidance. Self-Assembly at All Scales 29 orders of magnitude, same underlying mechanism! Carbon nanotube, about 10-8 m Our galaxy, the Milky Way, about m

58 What is NOT Self-Assembly? Planned and externally guided constructions Unstructured aggregates Living beings, because they are not made of pre-existing units Gazprom building, St- Petersburg, 10 2 m Sequential, limited scalability (from a few millimeters up to a few kilometers) Boxelder bug aggregate, 10-3 m Aggregation may serve a purpose (e.g., protection, collective decision- i making), but the structure itself does not. Various animals Self-assembly may have played a key role in the initial iti appearance of life.

59 Self-Assembly at All Scales A.Crystal structure of a ribosome [Yusupov et al., 2001] B.Scaffolded DNA origami of 100 nm in diameter [Rothemund, Caltech] C.Capillary SA of 10-μm-sized hexagonal plates into large crystal structures [Whitesides, Harvard] D. SA of 1-mm-sized hexagonal plates using capillary bond [Whitesides, Harvard] E. SA of electrical networks [Whitesides, Harvard] F. Electromechanical programmable parts specifically designed for SA [Klavins, University of Washington] G. A swarm of mobile robots capable of SA [Dorigo, UniversitéLibre de Bruxelles] 59

60 Ingredients and Features Ingredients of Self-Assembly: 1. Positive feedback: binding force, structural reinforcement 2. Negative feedback: repulsive li force (otherwise, (th merely aggregation) 3. Multiple reversible interactions 4. Randomness Features of Self-Assembly: Robustness: self-assembly deals with component defects, environmental noise and external perturbations Parallelism: self-assembly occurs wherever there are interacting components Scalability: self-assembly is a strategy applicable at all scales, and even across scales Ubiquity: structures form spontaneously at all scales in Nature, and most of them exhibit similarities across scales

61 Mti Motivation for Studying Stdi Self-Assembly SlfA (1) Scientific understanding of a broad variety of phenomena: Water ripples ~10-3 m Seashell ~10-2 m School of fish ~10 1 m DNA origami Tornado Cyclone ~10-7 m ~10 2 m ~10 6 m

62 Motivation for Studying Self-Assembly (2) Manufacturing and coordination of ultra-small distributed intelligent systems: Left: A microrobot (10 μm) is inoculating a red blood cellagainst some Left: A microrobot (10 μm) is inoculating a red blood cellagainst some kind of disease. Right: This nanoscaled device (1 nm) enables a nanorobot to catch HIV individuals and destroy them.

63 The quest for miniaturization How to go from the Alice robot (2cm in size)...to fully-fledged microrobots?

64 Limitations of the Top-Down Approach HOW TO ASSEMBLE THESE INTO THESE USING THESE? ANSWER: LET THEM ASSEMBLE THEMSELVES!

65 Self-Assembly of Micro-Systems Binding surface Non-binding surface Self-Assembly Hydrophobic interaction and capillary forces (in fluid) Van der Waals forces (in dry state) Experimental challenges: imaging (3D, high speed camera, microscope) and controllability 65

66 Intelligent and less intelligent agents Alice mobile robot MEMS 2cm 100µm Typical size: 2 centimeters Typical swarm size:< 100 units Sensing, computation, communication (only local) Controllable self-locomoted units (but quite noisy) Typical size: 50 to 500 μm Typical swarm size:> 10 4 units No sensing, computation or communication, but local interactions only Stochastic motion only

67 Multi-Level Modeling Methodology Macroscopic 1: rate equations, mean field approach, whole population Macroscopic 2: Chemical Reaction Network, stochastic ti simulations Microscopic 1: Monte Carlo model, 1 robot = 1 agent, non-spatial Abs straction Common metrics Microscopic 2: Agent-Based model, 1 robot = 1 agent, spatial Realistic: faithful representation of the robots (e.g., geometry, S&A) and the environment (e.g., friction, gravity, inertia) Physical reality: real experiments, imaging with overhead cameras, tracking software rimental ti ime Expe

68 Multi-Level Modeling Methodology Macroscopic 1: rate equations, mean field approach, whole population Macroscopic 2: Chemical Reaction Network, stochastic ti simulations Microscopic 1: Monte Carlo model, 1 robot = 1 agent, non-spatial Abs straction Common metrics Microscopic 2: Agent-Based model, 1 robot = 1 agent, spatial Underlying physical reality (and, therefore submicroscopic models) may vary dramatically! rimental ti ime Expe

69 Goal Ex.1: Self-Assembly of Alice Robots Formation of chains Physical testbedtb dfor models at multiple abstraction levels Self-assembly soft binding force (implemented through local IR communication) Advantages Controllable, but still miniaturized 19 robots, accelerated 16x Experimental ease: imaging, controllability, flexibility [Evans et al., ICRA 2010]

70 Single Robot Algorithm Timer expires OR Unaligned object detected Tumble Timer expires Start Wander Aligned object detected Bond Avoid Timer expires No recent communication OR No recently aligned object Deterministic vs. stochastic variant

71 Deterministic Controller Bonded Alices communicate with each other to determine the length of their chain msg = 4 msg = 1 msg = msg + 1 msg = 2 msg = 4

72 Deterministic Controller - Sample Results with 19 robots Submicroscopic (Webots) Macroscopic (stochastic CRN)

73 Stochastic Controller Every second, an Alice leaves its current bond depending on some probability. bilit

74 Stochastic Controller - Sample Results with 19 robots Submicroscopic (Webots) Macroscopic (stochastic CRN)

75 Ex. 2: Self-Assembly Micro-Systems Ground truth: spatial microscopic model - implemented in Netlogo Abstraction: shape of the building blocks, most physical effects (e.g., friction, fluid dynamics) Assumption: building blocks can be represented as simple agents with a certain orientation and a single preferential direction Discretization of alignment state Random motion is described by a Langevin equation (Brownian motion): friction term agitation term Collisions handled in a deterministic fashion. [Mermoud et al., AAMAS 2009]

76 Microscopic Non-Spatial Model Abstraction: spatiality, i.e. the position and orientation of each agent Assumption:the system is well-mixed, homogeneous and isotropic This Monte Carlo model does not capture spatiality, but still keeps track of each aggregation/disaggregation event in the system. Geometric approximation for encountering probability (volume swept): Energy-based formulation of the break-up probabilities, which depend on the alignment of the building blocks: Energy in current state si Misalignment System agitation Maximal bond energy

77 Macroscopic Model Abstraction: ti the state (i.e., whether h it is in an aggregate or not, and with which agent) of each individual agent Assumption: the system can be modeled using a mean field approach based on rate equations. Our model is based on difference equations that keep track of the number of individuals in each state. Capturing alignment by discretizing of the space of pairs into K different alignment states, each with a different break-up probability: Pair of building blocks in a given state i Number of single building blocks

78 Macroscopic Model Validation Excellent agreement between non-spatial models (red and blue). Spatial model (agent-based, in green) converges slower: lower mixing.

79 Mean field approximation validity N0 = 500 N0 = 100 N0 = 50 Also, the faithfulness of macroscopic model with respect to microscopic spatial model degrades as the total number of building blocks N0 decreases.

80 Optimizing Self-Assembly Agitation brings mobility to the building blocks, but also mitigates the stability of bonds: find the appropriate tradeoff! Non-linearity, exploration vs exploitation tradeoff: self-assembly is a difficult control problem! Optimization of control and design parameters requires accurate and computationally inexpensive models.

81 Conclusion

82 Take Home Messages Both types of stigmergy (quantitative and qualitative) play a crucial role in natural distributed building processes Not only self-organization and stigmergy but also environmental templates can be used for coordinating the activities of distributed intelligent systems Once again microscopic and macroscopic models can help in understanding, designing, and optimizing real distributed building systems

83 Take Home Messages Self-assembly implies the constitution of structures and ordered spatial pattern (rather than just aggregate) and constituting elements that incorporate all the required assembly mechanisms (although they are not necessarily endowed with self-locomotion capabilities) Self-assembly can happen at all scales and is a powerful, parallel, potentially cheap coordination principle for a number of manufacturing applications in which traditional techniques cannot tbe used or have an excessive cost

84 Additional Literature Week 12 Papers Bonabeau E., Guérin S., Snyers D., Kuntz P. and Theraulaz G., Three-Dimensional Architectures grown by Simple Stigmergic agents. Biosystems, 56(1): 13-32, R. Gross, M. Bonani, F. Mondada, and M. Dorigo. Autonomous self- assembly in swarm-bots. IEEE Trans. on Robotics, 22(6): , 2006.