Evolutionary Task Force

Included page "clone:symbrion-ec" does not exist (create it now)

Minutes for meeting in Stuttgart (2012-1-11) - 12 Jan 2012 16:03

Reader's guide

  • Search for string "[!]" in all the text to quickly jump to IMPORTANT elements (e.g. discussion, todo, open questions, etc.)
  • Organization: (1) summary of decision (2) exhaustive notes
  • [DECISION] tag marks ''real'' decisions. The rest is important information.

MORNING DECISIONS

Sub-task 2 (organism controller)

  • Yaoyao has to provide same measurements.
    • [!] should be done within one week

Sub-task 3 (internal rewards)

  • [DECISION]: Stick with distance for the demo

Sub-task 4 (simulation)

sub-task 4.1 (libor)

Threaded implementation
  • [!] Libor organizes a test check whether sensor updates are synchronized
  • Motivation: threaded controllers speed up the simulation (if only a little).
IR sensors: Currently a binary output with respect to a predefined threshold distance
  • [!] make it proportional (integer values)
Docking:
  • [!] Info: we agree on assuming that it is 2D planar docking
  • [!] Info: both cells need to agree to be docked to perform docking, no protocol yet. currently: all single robots are always ready and claiming they want to connect. Ie. two close candidates will connect.
    • TODO: Discussion Libor/Vojta
Requirement wrt. evolutionary runs (target: to render a video for demo)

* [!] we have to evaluate the requirements in terms of number of evaluation wrt. evolution.
* [!] ??WHO/WHEN??

Integration of energy simulation

* [!] [DECISION]: assume linear decrease of energy
* Goal: over-simplify (we dont need precise modeling for now),

Running the simulator on a server
  • We cannot go without GPU as OpenSceneGraph strongly relies on it. Too costly to re-program.
    • [!] [DECISION]: We have to go with a GPU machine
Sub-task 4.2 (Jean-Marc)
  • Exchange of genome btw organisms:
    • [!] [DECISION] : Lamarckism (but homog/heterogeneous is just a feature wrt. implementation - ie. it's your pb)
  • Decision: who is responsible within other subtasks (for implementation)
    • [!] [DECISION] : Task 1 will nominate someone by monday 2pm.
    • Task 3 is already done
    • Task 2 is Jean-marc
    • Berend is available for some coding if needed
  • [!] [DECISION]: Will use Stuttgart SVN (everything is already there), not launchpad

AFTERNOON DECISIONS

Global discussions and decisions

Agreement on organization

  • (1) [Swarm including egg] —(2)> [Organism] --(3)——> [Moving organism]
    1. Swarm incl. egg
      • Recombination, variation of body shapes
    2. Morphogenesis:
      • Explicit_1 (Wenguo/Christopher)
      • Explicit_2 (YaoYao.1)
      • Implicit_1 (Ronny)
      • Implicit_2 (YaoYao.2)
      • Implicit_3 (Michèle)
    3. towards moving organisms
      • AHHS (Graz)
      • GRN (Yaoyao)
      • CPG (Jean-Marc/Evert/Florian)
  • Also, can be reformulated as a cycle:
    • birth => learn to walk => (foraging) => reproduce
    • We do not do foraging in the demo

Questions

  • what will be in the report ? what will be presented as a demo ? outcome as paper(s) ?
  • [!] PROPOSITION / DECISION:
    • Morphogenesis:
        • DECISION: two methods (1 explicit, 1 implicit)
        • DECISION: the explicit is Wenguo/Christopher
        • DECISION: decision for the implicit option will be taken by the sub-task people and is due on monday 2PM "[!]"
          • Michèle will send a message reminding the indicators (deadline: tonight)
          • All are asked to fill the figures by friday afternoon for each implicit approaches
          • A skype meeting will be organized to choose which implicit approach is selected - friday afternoon
        • note: this is a pragmatic choice to target the demo based on what is available now.
        • note: all other approaches will be reported in the Report, incl. possible comparaison.
    • moving organism
        • AHHS/GRN/CPG are kept for the report, then decision will be made for the demo *just after the report*
        • decision for the demo could be one of them, or (part of) all of them (with meta-gene selection)
        • we will also do the combo version (meta-gene version)

Various Questions/Remarks:

Q: how to change the view in robot3D?
A: ''press the button'' in robot3D

Recommendations/Requirements:

Building blocks for morphogenesis:
  • Magic transportation (then true docking)
  • What is in the message for the recruitee when to arrive and with which orientation
  • [!] Berend is managing a dedicate System Description for the Evolutionary Task Force. (general source information)
  • everybody should contribute (what have you done, how it works)
  • Provide *detailed* description of experimental conditions.

=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=

=-=-= RAW-NOTES:BEGIN =-=-=
=-=-= RAW-NOTES:BEGIN =-=-=
=-=-= RAW-NOTES:BEGIN =-=-=

=-=-=-=-=

[Gusz presentation]

Past and present:
- mid-october: Evolution Cluster is acted, goal: integrated demo before the 4th review meeting
- 2-3 november: kick-off meeting, 4 sub-tasks, agreement on working protocol & targets (demo & joint papers). Planned weekly skype meeting (thursday morning)
- 11 january: review of progress, planning or research towards demo & joint publication(s)

Future:
- early march 2012: Annual progress report
- mid-april 2012: review meeting presentation
- Summer 2012: conference paper(s)
- Fall 2012: journal paper(s)
- Winter 2012/2013: big journal paper (Nature, Science, PNAS ?)

=-=-=-=-=

=-= SUB-TASK 1 : MORPHOGENESIS [Michele] =-=

General Introduction
- Agreed:
- viable shape
- number of robots [2-10]
- test for evolvability and controllability

Group 1: wenguo@UWE, christopher@UT
- explicit representation
- implemented in Robot3D, nested with Stage simulator ==> ? ==> robot3d is just for visualization
- demo with 25 robot - ?
- video shows *recruiting*, not evolution
- ie. this shows birth from egg to organism
- Discussion: Robot3D with nested Stage "[!]"
- ok for Swarm
- if swarm and organisms: must be extended, not that easy. (emulate 3D organisms as obstacle in Stage, running mixed controller in Robot3d (from Robot3d or from Stage)).
- note on robot3D: all features are there (incl. radio communication)
- evolution of Wenguo's representation
- Christopher did some tests, not with robot3d — preliminary tests (fitness: biggest one), sounds ok.
- pending: do it in robot3d

Group 2: Graz (Markus, Ronny)
- genotype is linear in the complexity of organisms (not size)
- not tested wrt. viability
- variation operators: point mutation, crossover, transposons (planned)
- evolved: target pattern, symmetries, etc.
- tested:
- netlogo, robot3d
- no recruiting for the moment, only focus on direct assembly (focus: recombination, mutation, explicit fitness function for shape)

Group 3: Ghent (yaoyao)
- inspired by genetic regulation
- two versions: centralized, decentralized
- variation: gene dupl., adaptive mut.
- properties: one-to-many, explicit
- assume: position/orientation sensing
- Discussion : how to dock with real robots? camera alignement? (ubisense will not be precise enough). Docking is not working in robot3D - how to do it?

Group 4: INRIA (michele)
- Stochastic Cellular Automaton
- implicit, stochastic, ration pheno/geno=1.2
- evolvability is fine
- limits: controllability
- tested on robot3D (teleport?)
- Q?: variability of shapes, controllability, efficiency of locomotion

SUMMARY:
- five different approaches
- explicit: wenguo, yaoyao.1
- implicit: yaoyao.2, michele, ronny
- assessing development:
- wenguo is 0 week away from robot3D (assuming stage is ok)
- graz is 1 week away from robot3D
- yaoyao is 2 weeks away from robot3D
- michele is 0 week away from robot3D (assuming magic teleportation is ok)

=-=-=

=-= SUB-TASK 2 : ORGANISM CONTROLLER [Juergen] =-=

task: control the behavior of the robot organism

General introduction:
- starting point:
- cpg (evert)
- cpg (florian)
- grn (yaoyao)
- ahhs (jurgen)
- proof of concept: integrate in robot3D
- Now:
- CPG (evert, florian, jean-marc)
- GRN (yaoyao)
- AHHS (jurgen)
- robot3d is used in experiment, approaches
Demo/Experiments:
- Arena: flat, surrounded by walls, 50x50 KIT-robot sizes
- organisms and eggs in arena
- organism pool: i, I, T, H
- each robot organisms operates on its own genome
- genomes are collected and transmitted all the time
- if an egg is in sensor-range, it is fertilized
- a genome embeds all three sub-genomes for the three competing method
- the choice of organism (i,I,T,H) is *fixed* per organism
- one genome value chooses among the three sub-genomes
Roadmap:
- scenario file incl. organism and eggs (Jean-marc, finished)
- communication org to egg (still open)
- communication btw ctls in the org (parly finished, all)
- teleporting of robots to egg for organism (still open)
- meta-controller with ctl-template (finished, Jean-Marc)

Group 1: Graz (juergen)
- AHHS
- online evolution (ie. evaluation of a solution starts from whereever the robot is), fitness: travelled distance (explicit)
- robust to noise, though there is no reevaluation.
- integration in robot3D, genome stored as one string
- still todo:
- collecting all genomes to write it to the eggs
- evaluation of the fitness seems corrupt (…?) — localization problem …?

Group 2: Ghent (Yaoyao)
- GRN approach (similarities with Floreano's AGE)
- first test with robot3d
- learns to walk in Robot3d (which morpho?)

Group 3: INRIA+STUTT+VU (Jean-Marc, Evert, Florian)
- oscillator + NN = sensory information to motor outputs
- IR-sensors[1-8] => [[NN]]=> amplitude, phase shift, frequenscy, angle offset => [[CPG]] => hinge angle
- weights of NN are optimized
- heterogeneous approach — one genome per robot
- cf. also homogeneous approach. (one genome per organism)
- no central clock, but start at the same moment (cheating in the simulation)
- results with heterogeneous approach
- running in robot3d
- mu+1 is used
- did it for snake with 6-modules
- "Q: homogeneous may be better, but need synchronization information (such as transmitting info from one block to the the other)"
- "Q: average distance travelled for the same organisms"

SUMMARY:
- three different approaches, overarching system including all these methods
- no obligation of choosing now
- all three approaches have been tested, Juergen's and Jean-Marc's also have experimental results.
- Yaoyao has to provide same measurements also. => one week "[!]"

=-=-=

=-= SUB-TASK 3 : INTERNAL REWARDS [Evert] =-=

task: internal rewards…

- Results with webots, open and closed arenas
- QI (gps and/or sensorimotor space), distance, learn, random —- compared on distance travelled within small intervals
=> for small intervals: QI, learn, random are more or less the same wrt. distance travelled. Far from distance.
=> for smaller intervals: similar.
(note: i'm not sure about this part, please refer to slides)
- used a ''GPS'' emulation with 10cm error
- QI is not that good because …? possibly because envt is not rich enough (but also relates to QI coverage wrt. epsilon value) "[!]"

=-=-=

=-= SUB-TASK 4 : SIMULATION [Libor] =-=

task: … implementation and speed issues

Group 1: Libor

- what's new in the simulator (since october @york)
- messaging system
- radiomessages: broadcast to all (zigbee) —- possible issue with bandwidth
- organism/messages: to physically connected neighbors
- controller/agent outfit
- from: threadless implementation
- to: threaded implementation (// performance)
- motivation: to solve a problem of synchronized sensor update — should tested further "[!]"Libor organizes a test "[!]"
- may improve the speed, "[!]" but remain to be tested "[!]" (ie. physics still runs on a single core, which is the main thing)
- IR sensors
- single beam irradiation pattern
- beam heading corresponds to robot body orientation
- Binary output with respect to a predefined threshold distance — "[!]" make it proportional (double or float value)
- Docking
=> Remark: based on position on the cells at the moment, plan for camera-driven docking with real robots "[!]"
=> "[!]" we agree on assuming that it is 2D planar docking
=> note: both cells need to agree to be docked to perform docking, no protocol yet. currently: all single robots are always ready and claiming they want to connect. Ie. two close candidates will connect. "[!]"
- LED not used (failing LED guidance system at the moment)
- Document online (symbrion-ec wiki)

- Speed issues
- close to real-time to 10 times slower.
- "[!]" we have to evaluate the requirements in terms of number of evaluation wrt. evolution.

- Simulation outlook
- support - priority #1
- integration of energy simulation
- "[!]" DECISION : assume linear decrease of energy —- goal: over-simplify (we dont need precise modeling)
- proposal : use time till last charge as indicator, implement it as a new sensor (battery level sensor), decrease linearly through time.
- server for final experiments - suggestions
- one server: 12 CPUs, 40GB RAM
- no video card available (no GUI, no GPU), remains questionnable.
- target headless simulator would be great but…
- "[!]" We cannot go without GPU as OpenSceneGraph strongly relies on it. To costly to re-program (and possibly, it would be slower). We have to go with a GPU machine.
- other pending tasks: on-board sim, sampling-based planning method for organism control/collision avoidance (lacking manpower due to priorities on simulation)

Group 2: Jean-Marc (with Berend, Vojta, Anne, Lutz) (talk done in the afternoon, reported here for consistency)

Done:
- meta-controller
- communication btw organisms, within organism
- CPG controller, heterogenous
- space for other controller

Todo:
- echange of genome btw organisms: once an organism is born, and then die, do we keep the evolved controller (lamarckism) or send the initial one (baldwin).
=> "[!]" Decision : Lamarckism (but homog/heterogeneous is just a feature wrt. implementation - ie. it's your pb)
- egg calling morphogenesis process

Questions:
- Who is responsible within other sub-tasks?
=> "[!]" …?
- Use launchpad?
=> "[!]" [DECISION]: will use Stuttgart SVN (everything is already there)

=-=-=-=

SUMMARY OF MORNING:
- many cross-talks, team spirit
- Gusz is happy, others agree

=-=-=-=-=-=

AFTERNOON SESSION

[Swarm including egg] —(1)> [Organism] --(2)——> [Moving organism]

(0) swarm incl. egg
- recombination, variation of body shapes
(1) morphogenesis:
- Explicit_1 (Wenguo/Christopher)
- Explicit_2 (YaoYao.1)
- Implicit_1 (Ronny)
- Implicit_2 (YaoYao.2)
- Implicit_3 (Michèle)
(2) towards moving organisms
- AHHS (Graz)
- GRN (Yaoyao)
- CPG (Jean-Marc/Evert/Florian)

Also, can be reformulated as a cycle:
- birth => learn to walk => (foraging) => reproduce
- we do the first two

Evolutionary Components
- representation (genome)
+ body
+ controller
- variation
+ mutation
+ crossover
- fitness def for selection

QUESTIONS
- what will be in the report ?
- what will be presented as a demo ?
- outcome as paper(s)

PROPOSITION / DECISION: ""[!]""
- AHHS/GRN/CPG are kept for the report, then decision will be made for the demo *just after the report*
- decision for the demo could be one of them, or (part of) all of them (with meta-gene selection)
- we will also do the combo version (meta-gene)
- morphogenesis:
- Decision: decision for ONE solution will be taken by the sub-task people and is due on monday 16th "[!]"
- note: this is a pragmatic choice to target the demo based on what is available now.
- note: all other approaches will be reported in the Report, incl. possible comparaison.

Report demo paper 1
(=>choice)
Wenguo/C ?
Yy.1 ?
Ronny ? ?
Yy.2 ?
Michele ?
AHHS +
GRN +
CPG +
Combo + ?

Open questions:
- trade-off btw favoring all in //, or arbitrarily selection some

Additional decisions of morphogenesis (target: report) "[!]" :
- Michèle will send a message within the next 3 days reminding the indicators
- All are asked to give feedback by monday "[!]"
- Then, all are asked to fill the table wrt. their approach and indicator

=-=-=-=-=

=-=-= RAW-NOTES:END =-=-=
=-=-= RAW-NOTES:END =-=-=
=-=-= RAW-NOTES:END =-=-=

** Additional notes from the coordinator intervention **

19/1: senior meeting
review meeting:
- 2-4th may, Stuttgart - confirmed?
- required presentation of the planned outcome
- mass production should have begin by the review meeting
- BUT some results from production should be available at the review meeting (4-5 people more, target 10-15 robots (5 each))
- target: mass-production is finished in june/july
- message:
- dont speak about prolongation
- simulation is good, but without hw support, will not be accepted
- joint ownership for the robots (agreement will be proposed)
- no shifting of the budget.
- distribution will be decided in great assembly

end.
Tags: | Comments: 0

Unless otherwise stated, the content of this page is licensed under Creative Commons Attribution-NonCommercial-ShareAlike 3.0 License