Progress report: week: 19.Dec - 08.Jan - 08 Jan 2012 18:32
General:
In the weeks 8 and 9 (19.Dec - 08.Jan) there was a skype meeting as the 20111221. The issue was to combine the implementation to get the approaches of sub-task 2 in a good piece of Robt3D-code. Therefore some decisions for details had to be made:
We decided on testing our approaches in a scenario which is similar to the scenario of the evolution cluster with skipping stuff done in sub-task 1. This means:
- there are organisms and eggs in the arena
- the organism have fixed shape for per experiment
- there is a super-controller which "evolves" (implicitly by the eggs) an integer in the range of 0-2. Each integer stands for one organism control approach. For testing, the range of the integer is 0-3 with CPG taking two values (one with random input, i.e., no sensor input).
- each agent of the organism operates on its own genome
- the genomes of all agents are stored together and transmitted "all the time"
- if an egg is in sensor-range, it is fertilized
The work to do for that and partly was done over the holidays (special thanks to Jean-Marc):
- scenario file including organism and eggs in Robot3D (jean-marc, finished 20120104)
- evolver of super-controller with the two CPG-variants implemented (as template) in Robot3D (as a skeleton) (jean-marc, 20120104)
- communication from organism to egg, including implementing communication range (berend, 20120110)
- communication between agents of the organism, including operating values during runtime on the one hand and the genome on the other hand (each approach on it's own, 20120110)
- teleporting = recruiting of other agents by the egg to build the robot (vojta or sub-task 1-people)
The code which is now provided by Jean-Marc is going to be used by the other approaches in the following week. The big switch ("super-controller") is implemented as a parameter at the moment, thus everybody can try their code.
Tags: | Comments: 0
Progress report: week 12.Dec - 18.Dec - 16 Dec 2011 14:35
General:
In week 12.Dec - 18.Dec the approaches were ported to the Robot3D simulator. Due to the fact that we seem to be the first and only really working with Robot3D a lot of smaller and bigger problems arose with the simulator. We do our best to improve the simulator that finally our approaches really can do their jobs.
We started to discuss detail about our subtask:
Shall every agent/module in the organism have its own evolution or shall
there be one single evolution for the whole organism?
Summy, so far:
every module is evolving by its own:
pro: code is nearly finished
pro: more easily ported to the real robots
con: in evo-cluster scenario: which module should fertilize the eggs which
are met during performance?
con: the evolution of one module depends too much on the situation of the
other modules
only one "evolution-manager" per organism:
pro: this is the case in our scenario
con: we (or wait for vojta to do so) would have to implement a new
compoment in the simulator
CPG controller (Evert, Florian, Jean-Marc):
Jean-Marc implemented the approach to Robot3D. Due to problems with the ode of the simulator (makes the simulator crash) there are no measurements, yet.
GRN controller (Yao):
I am testing my controller on the Robot3D in this week, the controller could receive the feedback of the actions and evolve its functional
roles in the organism so far. meanwhile I am learning to use the simulator and debugging.
AHHS controller (Juergen, Heiko):
This week porting AHHS to Robot3D was successful. There are still problems with the simulator, thus a test for the approach - in the meaning of distance walked after finishing online evolution - was not possible, yet.
Tags: | Comments: 0
CPG, Mu+1 and different bodies - 05 Dec 2011 10:41
We ran extensive comparisons of Mu+1 with CPG and different organism morphologies. The diagram below shows stats of distance travelled over 30 runs (with mu+1 settings as found with Bonesa) for four morphologies. Fitness was calculated with QI, based on gps, accelerometer and link force-feedback.
My conclusion: body shape certainly matters substantially and significantly (notches indicate 95% conf interval for median), but even poor bodies learn to move sufficiently, particularly in the confines of a walled arena, as indicated by the trace maps
Trace with four-legged body, note that the organism follows the wall (actually: tries to walk through it :-))
Trace with three-legged body (worst in the bar charts) - still acceptable coverage of arena area:
Tags: | Comments: 0
Progress report: week 28.Nov - 4.Dec - 02 Dec 2011 15:36
General:
In week 28.Nov - 4.Dec people of the sub-task started to implement their approaches in Robot3D. Due to bugs the progress is quite slow. Further, the two CGP approaches are going to be merged to one.
To make the decisions of last week (parameter's of e.g., the arena, for the comparison of the organism controller approaches) more visible, they are the basis of the method section of a pdf-file which is a draft of a paper. This document should be a living document. It is stored at the Symbrion repository (publications/Papers/paperOrganismControlComparison/). Currently it looks like this: paperOrganismControl.pdf
CPG controller (Evert, Florian):
GRN controller (Yao):
AHHS controller (Juergen, Heiko):
Since proof of concept (online evolution was shown last), the implementation of the controller to Robot3D started. Due to bugs in the simulated the progress is slow.
Tags: | Comments: 0
SUBTASK 2 Organism Control - 25 Nov 2011 14:20
General:
In week three (21.-27.Nov) we had a skype conference an Monday.
The minutes in short:
- one scene file for all controller tests:
arena: flat, 5x5 meters (50x50 KIT-robot sizes); implemented by Jean-Marc
- Organism pool: i (2 modules), I(8 modules), T (2+2+2 plus 1 middle module), H (4 modules for the horizontal bar and 2+2+2+2 modules for the "legs")
- detailed parameters (number of ticks/steps per evaluation, number of steps in total,…) are discussed later
- base class / Interface proposal or class for all controllers (provided by Evert) with a switch which is evolutionary controlled. The issue that maybe only one of the four controllers gets evolved (getting stuck in local optimum early) will be discussed next week.
CPG controller (Evert):
Proposed common controller interface, started analysis of CPG/MuPlusOne runs for varying morphologies. Initiated (with Jean Marc, who is visiting from Paris) port of CPG/MuPlusOne to robot3d and integration of sensory stimuli. Jean Marc is working on the arena and other infrastructure ('egg' mechanics in robot3d, etc).
CPG controller (Florian):
Our progress for the last week is to set up finally the simulation on several systems and we managed to integrate the controller into the simulation. My further steps are now to refine my algorithms and evaluate the performance according to the criterions stated in order to be comparable at the 15.12.2011.
GRN controller (Yao):
In this week, I am working on the duplication function of the GRN controller, meanwhile I also improved the structure of the genome. The work is focus on making more efficient feedback loops to genetic duplication. For the Robot3D, I already start to install it. ( My laptop OS is too old to the new version,so I am upgrading the OS to the support version at the moment)
AHHS controller (Juergen, Heiko):
We implemented AHHS in the FEP simulator and made first test runs in an open arena and in a maze. Fitness was measured by traveled distance. Fitness improvement was observed, thus first successful online evolution of AHHS was established. Further, we installed Robot3D on several machines in our lab. We still wait for an example controller within Robot3D and try to help with bug reports.
Tags: | Comments: 0
Subtask2 Organism Control - 18 Nov 2011 13:45
Progress Report Week Two
General:
In week two (13.-18.Nov) people worked on their approaches
and helped to debug the Robot3D simulator. In the weekly skype meeting
the comparison was postponed to the 15th of December with the
restriction that it is done in the Robot3D (if the simulator is
usable). A skype meeting was arranged for Monday next week.
CPG controller (Evert):
In addition to a very extensive set of runs where I combine mu+1 with
a central pattern generator in a 5-module organism, I ran a couple of
experiments in 4- and 3-module variants. In all cases, the organisms
locomote with this combination. Not always equally fast, but move they
do. Next step: walls around the arena and multiple organisms together.
CPG controller(Florian):
This week I invested a lot of time into the simulation and getting things
running. I pointed out some bugs (thread problems, gravity) and
incompatibilities (examples do not fit to trunk/devel). Meanwhile, I caused
several updates in the robot3d simulator and finally, I got it somehow running.
It is still not working fully for me, but it visualizes some robots.
GRN controller (YaoYao):
In this week, I am working on using GRN approach to organize the control
system in organism. In my simulation, GRN controller could detect the
position information of the robot and then GRN controller will indicate
the role of the robot in organism.According the role of the robot in
organism, robots will receive the different inputs from environment and
implement different functions. The next step will be using GRNs to
coordinate the actions of different robot (with different roles) in
organism.
AHHS controller (Juergen, Heiko):
On the one hand we changed the implementation of the evolution manager
of our approach to be able to do online evolution. First test runs
were made in our FEP simulator, without any analysis, yet. On the
other hand we tried to install the newest version of Robot3D and did
bug reporting to help the simulation team with their efforts. For
using Robot3D for the tests, a lot of work is still to do for the
simulation team.
Tags: | Comments: 0
SUB-TASK FORCE #2 Organism Control: Implementation Plan - 11 Nov 2011 15:20
Goal
General: Over the next three weeks, we'll be testing and comparing our
controllers: Juergen will perform trials with the AHHS controller,
Florian with his CPG controller, Yao with his GRN controller and Evert
with CPG controller.
For now, distance travelled per timespan is used as fitness measure and
the only goal is to have the organisms walk as long a trail as possible,
which may be circuitous. This is not the same as distance from origin.
who: Evert, Florian, Yao-Yao, Anne(helps with robot3d), Jürgen
For each controller:
- on-line, on-board adaptation/evolution
The controllers must adapt after the organism shape has formed to fit the controller to the body
shape. This must be done autonomously by the robots themselves.
"On-board" depends on the used simulator platform (which is variable
till 30. Nov.): some will be equipped of "self-location" of robots,
others not. As long as we are allowed to use any simulator - on-line
should be enough to be shown.
- Robustness
The shape of the organism is not, and cannot be, known a
priori. The ideal controller produces locomotion for any arbitrary
shape. Suggestion: single module, I, X and H-like?
- Reactive
The controller should be able to take sensory input into
account, for instance to develop obstacle avoidance behaviour. Starting
the tests with an arena surrounded with walls.
- Effective
Organisms must locmote using the controller.
Workplan of single approaches
CPG controller (Evert)
- with a number of body shapes, irregular as well as symmetric
- in an arena as we will have in the experiment: with walls, single
modules and other organisms as obstacles.
CPG controller(Florian)
First step is to get the simulation running with the current
implementation of the CPG. The next step is to validate the performance of
theCPG algorithms and evaluate according to the criterions stated. We
will use straight locomotion of several fixed organism structures and
measure the moved distance as fitness.
GRN controller (YaoYao)
The first step is testing the GRN controller with a number of shapes of organisms
in a simplified simulation and then getting the official Robot3D simulator
running (need some helps from Anne). After that, we plan to design an exploration
scenario and tested the scenario on the Robot3D. In the scenario,
the exploration range and the time factor will be considered as a fitness to organisms.
AHHS controller (Juergen)
The first step of the Graz group will be
to test the on-line evolution ability of the AHHS approach. This will
be done in an abstract and simple simulator called FEP. This first
test of feasibility study will even be conducted with single robot
modules with a fitness function of covered distance as it had been
agreed before.
During conduction of this first experiment we will additionally follow
two strategies: Getting the official Robot3D simulator running (with
help of Anne), so that the on-line evolution of AHHS can be tested
with robot organisms in the simulator. Also, after having finished the
feasibility study we will implement our scenario (using eggs and
"intrinsic fitness") within FEP.
Tags: sub-task-2 workplan | Comments: 0
