CCNR Khepera Lab

catch-up

2005-12-05T15:21:00.000Z

Hi Khep-people,

I agree with Edgar! Maybe it's time we meet again...
What do you think?

News

2005-11-09T14:39:00.000Z

Hi guys, after some time of silence in the Khep group, lets try to catch up...

MinGW (Minimalist GNU for windows) is available in blackstone so you can compile controllers just by using Makefiles.
Has anyone tried to write a document sumarising the documents (proposals, drafts, etc) we posted about the employment of Webots? (if not I can do it and post the first draft for corrections).

2005-10-19T14:00:00.000+01:00

For my PhD I need to run the following experiment, which tries to model the mating behavior of a certain frog species in which the male frogs evolved a signalling system. This prevents situations where a male unknowingly clasps another male (a situation with no evolutionary benefit). The clasped male lets the clasper know he's made a mistake.

Experimental setup:
- 3 identical mobile agents (but internally: 2 male, 1 female) with proximity sensors
- a 2D environment with energy sources
- female rewarded fitness for collecting energy sources
- males rewarded for being in close proximity to female

Additional fitness rules:
- if the two males are in close proximity neither get rewarded
- optionally, the female could also stop being rewarded for collecting food while being "clasped". While this has no direct biological counterpart it could pressure females to pick up on the male signals and use them to get rid of clasping males. The males would then have to evolve cheater detection abilities...

Note that no explicit communication channels are provided, i.e. the evolved communication has to be completely embodied.

It would be useful to test the experiment on real robots, especially because the evolved communication is going to be embodied. It will be interesting to see whether a richer (physical) environment automatically results in richer communication!

On birds, humans, and the morphodynamics of singing and dancing

2005-10-19T09:54:00.000+01:00

OK, guys. I'm interested in the evolution and ontogenesis of syntax.

Before you start thinking about human language, please think about birds. Some songbirds display non-stereotyped song structure. Some are better at doing this than others. For example, bengalese finches (Lonchura striata var. domestica) are much better that the non domesticated relatives white-back munia (Lonchura striata) :) ... sorry… these latin names, I just couldn’t resist the temptation … hope you enjoy them as much as I do. And gives a touch of good old natural sciences credibility. But I’m sure one day we will have a khepera-6 (khepera striata var. domestica)

Now, according to traditional views in linguistics (of the human sort), a protolanguage mainly constituted by “names of things” evolves then to a real language after the posterior evolution of a syntactical structure (cf. Bickerton, D.(1995), Language and Human Behavior, Seattle: University of Washington Press; cf. Pinker, S. (1994), The Language Instinct, New York, William Morrrow).
If you are sleeping here it's ok, but awake now! Here comes the the juicy part.

But in his work about bird songs – which he later hypothesizes being applicable to human language - Kazuo Okanoya challenges that assumption (cf. Okanoya, K., Sexual Display as a Syntactical Vehicle: The Evolution of Syntax in Birdsong and Human Language through Sexual Selection, in Ed. A. Wray (2002) The transition to Language, New York: Oxford U. Press). According to Okanoya, sexual selection for sequence complexity in bird songs, could result from serially ordered courtship display, involving voice – singing – and dancing, then originating a finite-state syntax. According to Okanoya hypothesis, semantics evolved separately, gradually introducing semantic tokens, initially with semantics very diluted in the flow of syntactical structure. These first vocalizations would have been of an aesthetic kind: a rhythmic pattern.

The kind of linguistic studies that we have found in Artificial Life so far (e.g. Luc Steels approach, see for instance the rationale behind the Talking Heads Experiment) draws on the assumption that a protolanguage mainly constituted by “names of things” must be in place if syntax is to evolve or ontogenetically develop.
I think it would be of great interest to setup an experiment in robotics whereby the conflicting (and more sensible in my opinion) Okanoya hypothesis is tested.

Another issue, for a different experiment, is treating body movement (and body articulation, namely limbs) as a condition for the evolution/development of complexity in sound output. (See, for instance, Leroi-Gourhan, A (1966) Le geste et la Parole, Paris :Albin Michel). Again, first and foremost at the syntactical/compositional level. Leroi-Gourhan is a paeleontologist but there is also evidence of this in neuroscience studies and developmental psychology.
Evolutionary robotics seems to be an appropriate tool to investigate in this area through the finding of existence proofs. Working with the khepera gripper can be in my horizon here.

These investigations require dealing with different stages of complexity of the experiments. On the other hand I see it as progressing from thought experiments to simulations and desirably to a physical robot.
Due its “functional plasticity” as a simulator, webots seems to be the right tool, as a very functional interface between thought experiments and physical robots, to pursue my research along these lines. Actually I see it as essential to my work right now and in the foreseeable future.

Just one last remark: a much broader conceptual/philosophical interest of mine is behind these investigations. That is the framework of morphodynamics and the possibility of physics (as opposed t a logic) of meaning, in the sense pursued by mathematicians like René Thom or Jean Petitot. Perhaps I must keep this to myself and to you guys J but just for the record…

contribution coming soon

2005-10-19T07:55:00.000+01:00

ooops, dificult day yesterday, didn't squash as thomas did :)
My contribution will be here in the next hour.

Our lab in all its beauty

2005-10-18T16:16:00.000+01:00

fig. 12: a khepera in its natural habitat.

2005-10-17T22:39:00.000+01:00

hmm... it's late, squash training sucked once more and i'm hungry. nonetheless, i'll try to add some arguments for how a khepera-lab and particularly the webots simulator could be useful.

the lab could encourage group work in a way that's open to anyone interested in robots. seems nothing special, but it's definitely missing from this institution. there is group work of course but only between faculty members and in a very secretive way. specifically, it can become useful as a teaching tool for EASy students who want to get their hands on either i) a real robot, ii) a working(!), reliable simulator or iii) both. I can imagine different forms of competitions as part of e.g. the adaptive systems course: beer-style block pushing, simplified soccer (e.g. penalty shoot out), sumo, wall avoidance races, you name it... now, regarding webots: there will always be shortage of robots if they were used as a teaching tool. but given enough floating licences, students could start in simulation and then take turns in transfering their results... there is always the argument of building your own simulator of course. however: i) code exists to be re-used ii) the EASy course accepts students not very proficient in programming iii) a common code base makes sharing as well as comparing controllers/results easier iv) 3d, physics, image processing etc. is hard for everybody, but provided by webots v) probably many more reasons here, please extend...
still group work: at epfl (for example) several big projects exist revolving around a rather general platform. one example is ijspeerts "salamander" project. students, ph.ds, postdocs and faculty take on different parts of research more or less related to the general theme. thereby i) new projects are being kickstarted by having something to start from, ii) people can exchange ideas on related stuff which can be more efficient and definitely more fun, iii) it looks cool to have a big project that you can show off rather than many small projects only iv) more here, if my brain was awake ...
CCNR: CCN and ROBOTICS(?)
think eduardo might have mentioned this already, but when presenting results in a paper it seems to add to your credibility if in addition to simulation your concept also works on a real robot. don't know if that's good or bad, but if by having this group and sufficient resources we can help people implementing their ideas in the real world that seems like a good idea. generally, not only would having this group, the robots and simulators benefit members of the group itself but we could also function as tutors/instructors/helpful people for others once we know how to do stuff right...
...

there was more, i'm sure, when we were having the roundtable. have to catch my train though, so maybe i'll add more tomorrow...

eduardo's (under construction)

2005-10-17T19:05:00.000+01:00

My interest in the khepera lab is manifold. Firstly, I am interested in taking the ideas that I am developing as part of my DPhil into the physical realm as (a) a proof of concept and (b) to put into practice the idea that the embodiment and situatedness will make it 'easier' on the requirements of the controller. Secondly, I am interested in it as a common platform where collaboration among other members of the lab can take place. And thirdly, I am interested in it as a platform where to get hands-on experience with robots.
Which brings me to a very important question: Why kheperas?
(a) Kheperas are arguably the simplest robot platform; (b) certain aspects of the khepera and simple environments can be easily simulated; (c) a number of evolutionary robotics work is based on this same architecture and (d) there's a good tool for simulating the more complex add-ons to the khephera as well as more complex/realistic environments.
Which projects?
1) Evolve a khepera robot controlled by a (non-modified-) CTRNN to do first phototaxis on a couple of lights and then path integration using only the infrared sensors.
2) Evolve a khepera robot controlled by a (non-modified-) CTRNN to do perform a task inspired on parental impriting in birds. A circumference of a particular size is shown during the first stages of the khepera's run. This object he must learnt and the khepera must be able to recognise it by its size later on during its life (see [1] for more details on the task).
The interest in (1) and (2) is mainly in the analyses of the internal dynamics coupled with the interaction of the agent and its environment so as to produce learning and memory behaviour.
3) Slightly more far-fetched is the last one: phototaxis by a blind but moving khepera, aided by the sounds produced by its non-blind-mate khepera which is stationary (can only rotate on the spot) . The purpose of this task is to study the communication that evolves. Is it a referential language (i.e. the light is on the left corner of the box we are in)? or perhaps a relational language (i.e. move to your right, hot, hotter,.. nooo to your left, almost there!)?

Report: My research and the use of a simulator

2005-10-17T11:50:00.000+01:00

I am interested in neural models for visual object recognition and Evolutionary Robotics. In particular, I want to explore the use of active vision and attentional mechanisms for visual object recognition in a robot. The next stage in my research focuses in the interaction between the robot and its environment in visually guided tasks. By using simple neural models and an active vision approach in an agent, I will study the exploitation of attentional mechanisms in visual object recognition tasks and the possible interaction between actuators and the visual system and between the navigational strategies and the visual system.

The use of a robot simulator is very important to this kind of work because of the flexibility and controlled way to reproduce realistic conditions that this tool offers. Webots [1] is a good option as a simulator because is a well tested and documented tool. Also, it does not imply spending time in the implementation of a simulator with complex requirements (real cameras, 3-D modeling, realistic illumination, etc).

[1] Michel, O./Cyberbotics Ltd (2004). WebotsTM: Professional Mobile Robot Simulation, pp. 39-42, Int. Journal of Advanced Robotic Systems, Vol. 1 Number 1 (2004), ISBN 1729-8806.

mailing list

2005-10-15T13:07:00.000+01:00

the mailing list for the group has been set up: kheplab@sussex.ac.uk

Patricia's report: Objectives and suggested Webots comments

2005-10-14T16:09:00.000+01:00

Patricia Vargas (*)

My primary aim is to develop and evolve robot controllers (for simulated and real robots) to assess two principal artificial bio-inspired systems that are currently under preliminary investigation:

1) GasNet models

The Study of GasNet models encompasses the implementation of a novel idea of gas dispersion: a non-isotropic model. Additionally, other GasNet architectures, among which are 1D (one dimension), 3D (three dimensions) and point to point (i.e., with non-spatial relation) will be investigated in conjunction with the non-isotropic model.

2) AHS

AHS is the acronym for Artificial Homeostatic System. This system is an artificial model developed during the PhD studies of the researcher, back in Brazil (Vargas, 2005a; Vargas, 2005b). This model is comprised of artificial neural networks and an artificial endocrine system. The proposal is to implement a process, similar to the GasNet model of diffusion of gases, to control the influence of artificial hormones over the artificial neural networks.

Regarding Webots software:

- A copy of the Webots software simulator (Michel, 2004) is currently under evaluation in CCNR labs.

- According to their creators, this software provides a rapid prototyping environment for modelling, programming and simulating mobile robots (both 2D and 3D simulations) in JAVA and C++ language.

- It also has robot libraries that enable the transfer of the controller program to many commercially available real mobile robots.

- One may argue that a powerful, confident and robust simulator like Webots might at least corroborate the validation of the approaches developed by the “Khe-Group”. In addition, due to its many features, the benefits of the use of this simulator for the research group seem to be substantial. For instance, all research members of CCNR might benefit from its use. As a result, the continuous use of the software will prove valuable for it prevents each researcher from spending a considerable amount of time developing its own specific simulator. All things considered, we strong recommend the acquisition of this particular software.

(*) part of this text is present at the SECSE-Report number C201.

References

Michel, O./Cyberbotics Ltd (2004). “WebotsTM: Professional Mobile Robot Simulation”, pp. 39-42, Int. Journal of Advanced Robotic Systems, Vol. 1 Number 1 (2004), ISSN 1729-8806

Vargas, P. A. (2005a). Bio-Inspired Computing Systems: Synthesis and Application in Computational Intelligence and Artificial Homeostasis, PhD Thesis Dissertation, UNICAMP, Campinas, Sao Paulo, Brazil.

Vargas, P. A., Moioli, R., de Castro, L.N., Timmis, J., Neal, M. & Von Zuben, F. J. (2005b). “Artificial Homeostatic System: a Novel Approach”, Proc. of the ECAL'2005, University of Kent, Canterbury, Kent

, Sept. 2005, 754-764.

khep-lab meeting

2005-10-10T19:12:00.000+01:00

several of us think that it would be appropriate to meet and discuss our research projects in relation to khep-lab, what we are doing, what we are interested in, how we can join efforts, etc. let me know if you would be interested in having such a meeting. one option could be friday 1pm-2pm. ??

Log Off instead

2005-10-07T14:32:00.000+01:00

Hey Kheguys, when remote desktop connecting to blackstone, make sure that to end the session you use the "log off" (as opposed to the "disconnect") button. Otherwise other khepeople will not be able to log on.
cheers!
Fernando

remember to log off

2005-10-02T09:32:00.000+01:00

from blackstone as soon as you stop using it...

full webots license

2005-09-29T15:02:00.000+01:00

the webots licence file has been installed. should run without restrictions now. mr. bond is also setting up remote desktop logins for the current members of the blog.

new focus and categories

2005-09-29T12:09:00.000+01:00

as a result of our wanting to encourage group projects and collaborations there was a transition from custom simulations and robosoccer to a more open focus on khepera-like robot experiments and simulations. at the moment our efforts are targeted at setting up a workspace for the group including several webots licenses, kheperas and their arena.

_technical:
The name of the computer where the group's first licensed Webots is installed as well as sitting next to the arena is: blackstone.rn.informatics.scitech.susx.ac.uk
The licensed node-locked webots version as well as the option to remotely log on should be available very soon! The use of this computer should be partially temporal. By which we mean one person should probably not be logged on for an entire week... unless people agree it's for a good cause. This rule should encourage us as the group to get a number of floating licenses. This is where good cause comes in: e.g. producing preliminary results using webots could help us convince people to buy more stuff.

_ettiquete:
1. using small fonts we get more posts on a page

2. suggested category hack: use underscores to tag keywords that you want to use as a category, e.g. the string _simulation anywhere in the post should be found with a blog search for "_simulation". later we can hopefully add permanent links to search for particular tags.

3. use real names for you user account

more things to try

2005-06-01T15:26:00.000+01:00

at the moment it seems like it's trivial to evolve goal seeking behavior, while approaching the ball is almost impossible. the only differences being the color (1,0,0) vs. (1,1,1), and the size. to check if this is what makes it so much harder we're currently evolving with a white ball against all gray walls and goals, as well as variable angles between sensory rays.

other things we could try:
1) subsumption architecture;
2) having same weights from all sensors, i.e. treating all sensors the same, reducing number of parameters for sensors to a fifth.

small first step

2005-05-30T00:03:00.000+01:00

success has been achieved when evolving an agent to go towards one of the goals (the white one) starting from a random position and orientation in the field (using the pessux funkadelic version with the set up described bellow). the fitness function used was simply the distance from the agent to the desired goal after a predetermined time limit relative to the starting distance, more formally: f=1-fd/sd where fd is the distance at the last time step and sd is the distance at the start of the trial. 10 trials are performed and the average is taken. additionally, if f is negative then fitness is set to 0 for that trial. more about the evolutionary algorithm used to come... at the moment we are trying to obtain a similar result now moving towards the ball (red) placed randomly in the field.

on the agent's controllers

2005-05-29T23:58:00.000+01:00

we are using a three layer network, sensory input (20 nodes) 5 sensory rays with 4 nodes per ray, one for the distance and the other three for the color (ie RGB), 20 nodes in total. The sensory node is bilaterally-symmetrically connected to the hidden layer, containing 6 fully interconnected nodes. These are, in turn, connected to 3 output nodes, one controlling the forward speed, and the other two competitevely controlling the angle rotation of the agent. The bilateral symmetry in the weights from the sensory to hidden layer was added to reduce the number of parameters. At the same time, the rest of the network was made non-symmetric in order to encourage non-reactive behaviour as well as initiative from the agent. Preliminary experiments using all-symmetric controllers lead to purely reactive agents. Regarding the output and movement of the agent, the use of three output nodes (as opposed to the commonly used 2 nodes, each controlling the velocity of each lateral wheel) promises to be very succesfull for mainly two reasons: 1) agents can move forward or stop completely independent of any side movements (as opposed to the commonly observed eliptic trajectories towards the goal) and 2) agents control the turning with opposing redundant nodes. more to come, and as always - suggestions welcome.

pro evolutionary soccer light coming soon

2005-05-29T19:49:00.000+01:00

Despite the overwhelming successes in the development of the simulation based on funkalistic approaches to pro evolutionary soccer already three versions ahead, work on a light version is soon to be finalised. The light version is a minimalistic C implementation of the agents, ball and soccer field with the simplest physics possible. No comparisons between speed have yet been made, but hopefully some computation will be saved in the longer run, as well as resolving some issues in interoperability. Hope you enjoy it soon when it comes out!!! Check out this very rough preview video.

version 0.03 ...

2005-05-27T11:58:00.000+01:00

evolution has begun. the current setup: our agent has to identify the red ball against walls and goals of different colors. he has to move towards it and ideally "carry" or "kick" it in the direction of the white goal (whose side is randomly changing).

what you see is the agent on the football pitch, his neural network (5 nodes per sensor ray for distance and rgb values), the distributed genetic algorithm and real-time plot of the agent's performance...

dribbling

2005-05-26T18:11:00.000+01:00

a nice idea by eric: an easily implemented way of enabling agents to "dribble" is to simulate a form of magnetism between agent and ball. basically, when close to the ball an agent could turn on a weak magnet that pulls the ball towards its body (simply addForce() in ODE) and whenever it wants to "kick", it pushes the ball away from it's body. since we're not talking real physics anyway, this could allow for interesting strategies, especially in competitive co-evolution...

decomposing indoor-soccer playing for the fitness function

2005-05-25T11:00:00.000+01:00

search for the ball, move to it (ie. balltaxis)
search for the goal, move to it
chase another agent (?) (might come in handy at some point)
dribble with ball from point a to point b
shoot and score!
goal keeping skills, stop the ball from going in... (similar to Beer's circle catching agents)

more..

on our model's arbitrary level of complexity

2005-05-25T10:53:00.000+01:00

the project can be expanded in many directions:

from one agent per team to n
from fixed sensors to evolving sensor positions/angles
from static sensors to dynamically moving sensors (online)

and its focus can also be tuned to people's interest:

co-evolution
incremental evolution for getting agents to do more than `just' one thing
evolution of sensors / morphologies
evolution of communication

the idea is to start off as minimal as possible and add as we go along.

Some preliminary specifications (to be expanded)

2005-05-23T15:33:00.000+01:00

1) The pitch: robots will play INDOOR soccer, meaning there is no boundaries. Lines are replaced with walls. 2) Sensors: n sensors returning the distance and "color" value (rgb) of an intersecting object. 3) Objects: walls (black), goals (different colors), robots (blue and green), ball (red). 4) Dimensions: pitch will be according to robocup specifications, see here. Robot dimensions will be published once agreed on.

first steps

2005-05-23T12:31:00.000+01:00

pro evolutionary soccer v0.01 created this weekend:

still tuning friction parameters and stability...

about pessux

2005-05-20T21:30:00.000+01:00

it all started one late friday night (20th of May)...
pessux stands for pro evolutionary soccer at sussex and it's pronounced more or less like this:
psks..

el jefe and chebot