Loading ...
Sorry, an error occurred while loading the content.

New paper “HyperNEAT for Locomotion Control in Modular Robots”

Expand Messages
  • evertwh2004
    We are very pleased to announce our new publication HyperNEAT for Locomotion Control in Modular Robots, which will appear in the proceedings of the 9th
    Message 1 of 6 , Jun 4, 2010
    • 0 Attachment
      We are very pleased to announce our new publication "HyperNEAT for Locomotion Control in Modular Robots," which will appear in the proceedings of the 9th International Conference on Evolvable Systems (ICES 2010). The manuscript is available here:

      http://www.few.vu.nl/~ehaasdi/papers/Modular-Locomotion.pdf

      In this paper we introduce the idea of modular differentiation with HyperNEAT: we use a HyperNEAT encoding to generate homogeneous (in the literal sense) but varying neural network controllers for individual robot modules that are linked together into a larger `organism.' These controllers allow the multi-robot to locomote purposefully and even somewhat reactively.
      This paper constitutes a proof of concept and we hope to expand on this work to allow HyperNEAT-based control in arbitrary and ultimately developing multi-robot organisms.

      We would like to thank Jeff Clune for pointing HyperNEAT out to us and everyone on this list - Ken Stanley especially - for the fruitful discussions.
    • Jeff Clune
      Hello Evert- Congratulations on this paper! I look forward to reading this, and am glad you are finding HyperNEAT appropriate for your challenge. Best regards,
      Message 2 of 6 , Jun 4, 2010
      • 0 Attachment
        Hello Evert-

        Congratulations on this paper!

        I look forward to reading this, and am glad you are finding HyperNEAT
        appropriate for your challenge.


        Best regards,
        Jeff Clune

        Digital Evolution Lab, Michigan State University
        jclune@...
        www.msu.edu/~jclune




        > From: evertwh2004 <evert@...>
        > Reply-To: "neat@yahoogroups.com" <neat@yahoogroups.com>
        > Date: Fri, 04 Jun 2010 15:00:34 -0000
        > To: "neat@yahoogroups.com" <neat@yahoogroups.com>
        > Subject: [neat] New paper „HyperNEAT for Locomotion Control in Modular Robots‰
        >
        > We are very pleased to announce our new publication "HyperNEAT for Locomotion
        > Control in Modular Robots," which will appear in the proceedings of the 9th
        > International Conference on Evolvable Systems (ICES 2010). The manuscript is
        > available here:
        >
        > http://www.few.vu.nl/~ehaasdi/papers/Modular-Locomotion.pdf
        >
        > In this paper we introduce the idea of modular differentiation with HyperNEAT:
        > we use a HyperNEAT encoding to generate homogeneous (in the literal sense) but
        > varying neural network controllers for individual robot modules that are
        > linked together into a larger `organism.' These controllers allow the
        > multi-robot to locomote purposefully and even somewhat reactively.
        > This paper constitutes a proof of concept and we hope to expand on this work
        > to allow HyperNEAT-based control in arbitrary and ultimately developing
        > multi-robot organisms.
        >
        > We would like to thank Jeff Clune for pointing HyperNEAT out to us and
        > everyone on this list - Ken Stanley especially - for the fruitful
        > discussions.
        >
      • Ken
        Evert, thank you for making this paper available. I just finished reading it. I like the idea of applying HyperNEAT to the individual modules of a
        Message 3 of 6 , Jun 4, 2010
        • 0 Attachment
          Evert, thank you for making this paper available. I just finished reading it. I like the idea of applying HyperNEAT to the individual
          modules of a multi-module robot. Inputting the module location into the CPPN (which is reminiscent of multiagent HyperNEAT) is a clever idea (which you call "modular differentiation") for creating robust robots of this type and it looks like your results came out well. I was wondering if you plan to make available any videos of the modular robot from this experiment? I'm looking forward to seeing your future work in this direction.

          ken

          --- In neat@yahoogroups.com, "evertwh2004" <evert@...> wrote:
          >
          > We are very pleased to announce our new publication "HyperNEAT for Locomotion Control in Modular Robots," which will appear in the proceedings of the 9th International Conference on Evolvable Systems (ICES 2010). The manuscript is available here:
          >
          > http://www.few.vu.nl/~ehaasdi/papers/Modular-Locomotion.pdf
          >
          > In this paper we introduce the idea of modular differentiation with HyperNEAT: we use a HyperNEAT encoding to generate homogeneous (in the literal sense) but varying neural network controllers for individual robot modules that are linked together into a larger `organism.' These controllers allow the multi-robot to locomote purposefully and even somewhat reactively.
          > This paper constitutes a proof of concept and we hope to expand on this work to allow HyperNEAT-based control in arbitrary and ultimately developing multi-robot organisms.
          >
          > We would like to thank Jeff Clune for pointing HyperNEAT out to us and everyone on this list - Ken Stanley especially - for the fruitful discussions.
          >
        • Jeff Clune
          Hey Evert- I just read the paper. It was very well-written and a fun read. Congrats on the results! I second Ken s interest in videos if you have them. I would
          Message 4 of 6 , Jun 19, 2010
          • 0 Attachment
            Hey Evert-

            I just read the paper. It was very well-written and a fun read. Congrats on
            the results!

            I second Ken's interest in videos if you have them. I would love to see what
            your gaits look like. It is hard to visualize them from the screen shots.

            Your paper prompted a few questions in my mind:

            1) Why did you use the 'habituation' model for inputs (where you only
            generate inputs values for new stimuli)? Did you only try this after the
            more traditional way of doing it (just feeding the raw range-finder inputs
            into the network) did not work?

            2) I really liked your idea of having a default behavior (controlled by the
            biases in the output layer), with the rest of the ANN only getting non-zero
            inputs when objects are sensed. Did you evolve the biases too, or hand code
            them in some way? If you evolved them, what method did you use (what were
            the coordinates of the source/from node fed into the CPPN when querying for
            a bias value)?

            3) You say that the fact that some of the lines are different in figure 9
            indicates that the ANN modules are different (in their wiring and function).
            But wouldn't you see different lines even with identical controllers in each
            module because different modules are getting different inputs (e.g. when one
            leg moves close to a wall)?

            4) Did you see any symmetry (e.g. left-right) or other form of regularity in
            the gaits (like I see in my HyperNEAT-evolved gaits)?

            Overall I loved the work and look forward to more. Thanks again for the kind
            acknowledgement.

            Best regards,
            Jeff Clune

            Digital Evolution Lab, Michigan State University
            jclune@...
            www.msu.edu/~jclune




            > From: evertwh2004 <evert@...>
            > Reply-To: "neat@yahoogroups.com" <neat@yahoogroups.com>
            > Date: Fri, 04 Jun 2010 15:00:34 -0000
            > To: "neat@yahoogroups.com" <neat@yahoogroups.com>
            > Subject: [neat] New paper „HyperNEAT for Locomotion Control in Modular Robots‰
            >
            > We are very pleased to announce our new publication "HyperNEAT for Locomotion
            > Control in Modular Robots," which will appear in the proceedings of the 9th
            > International Conference on Evolvable Systems (ICES 2010). The manuscript is
            > available here:
            >
            > http://www.few.vu.nl/~ehaasdi/papers/Modular-Locomotion.pdf
            >
            > In this paper we introduce the idea of modular differentiation with HyperNEAT:
            > we use a HyperNEAT encoding to generate homogeneous (in the literal sense) but
            > varying neural network controllers for individual robot modules that are
            > linked together into a larger `organism.' These controllers allow the
            > multi-robot to locomote purposefully and even somewhat reactively.
            > This paper constitutes a proof of concept and we hope to expand on this work
            > to allow HyperNEAT-based control in arbitrary and ultimately developing
            > multi-robot organisms.
            >
            > We would like to thank Jeff Clune for pointing HyperNEAT out to us and
            > everyone on this list - Ken Stanley especially - for the fruitful
            > discussions.
            >
          • Evert Haasdijk
            Hi Jeff, Thanks for the interest and kind words. It was a fun experiment and we are enthusiastically moving forward with it. I ll try and get some videos out
            Message 5 of 6 , Jun 20, 2010
            • 0 Attachment
              Hi Jeff,

              Thanks for the interest and kind words. It was a fun experiment and we are enthusiastically moving forward with it. I'll try and get some videos out this week.

              Your paper prompted a few questions in my mind:

              1) Why did you use the 'habituation' model for inputs (where you only
              generate inputs values for new stimuli)? Did you only try this after the
              more traditional way of doing it (just feeding the raw range-finder inputs
              into the network) did not work?

              As you suggest, it was the result of some trial-and-error; it is fairly straightforward to achieve steady -i.e., non-reactive- locomotion without any sensors. It proved much harder to get any reactivity (i.e., negotiating obstacles) into the behaviour. Using raw range-finder inputs didn't produce that, that led us to try the habituation model. That being said, we're still working towards more prominent obstacle avoidance/negotiation.


              2) I really liked your idea of having a default behavior (controlled by the
              biases in the output layer), with the rest of the ANN only getting non-zero
              inputs when objects are sensed. Did you evolve the biases too, or hand code
              them in some way? If you evolved them, what method did you use (what were
              the coordinates of the source/from node fed into the CPPN when querying for
              a bias value)?

              All the substrate weights were determined by the evolved CPPN. I'm not entirely sure about the technical detail of the biases, I hope Andrei (who's on this mailing list as well) will answer this one in more detail.


              3) You say that the fact that some of the lines are different in figure 9
              indicates that the ANN modules are different (in their wiring and function).
              But wouldn't you see different lines even with identical controllers in each
              module because different modules are getting different inputs (e.g. when one
              leg moves close to a wall)?

              The spikes in the graphs are caused by obstacles coming into or disappearing from sensor range. The base levels of the graphs (particularly different in 9c) differ only because of modular differentiation. I think, but can't be certain without further analysis, that the direction and magnitude of the spikes differs in part because of modular differentiation, as well.



              4) Did you see any symmetry (e.g. left-right) or other form of regularity in
              the gaits (like I see in my HyperNEAT-evolved gaits)?

              Yes we did: typically, after a certain  amount of evolution, we would see symmetrical movement between the left and right legs. Not at all something that we pressed for; it just emerged.

              Cheers,

              Evert


              Overall I loved the work and look forward to more. Thanks again for the kind
              acknowledgement.

              Best regards,
              Jeff Clune

              Digital Evolution Lab, Michigan State University
              jclune@...
              www.msu.edu/~jclune

              > From: evertwh2004 <evert@...>
              > Reply-To: "neat@yahoogroups.com" <neat@yahoogroups.com>
              > Date: Fri, 04 Jun 2010 15:00:34 -0000
              > To: "neat@yahoogroups.com" <neat@yahoogroups.com>
              > Subject: [neat] New paper „HyperNEAT for Locomotion Control in Modular Robots‰
              >
              > We are very pleased to announce our new publication "HyperNEAT for Locomotion
              > Control in Modular Robots," which will appear in the proceedings of the 9th
              > International Conference on Evolvable Systems (ICES 2010). The manuscript is
              > available here:
              >
              > http://www.few.vu.nl/~ehaasdi/papers/Modular-Locomotion.pdf
              >
              > In this paper we introduce the idea of modular differentiation with HyperNEAT:
              > we use a HyperNEAT encoding to generate homogeneous (in the literal sense) but
              > varying neural network controllers for individual robot modules that are
              > linked together into a larger `organism.' These controllers allow the
              > multi-robot to locomote purposefully and even somewhat reactively.
              > This paper constitutes a proof of concept and we hope to expand on this work
              > to allow HyperNEAT-based control in arbitrary and ultimately developing
              > multi-robot organisms.
              >
              > We would like to thank Jeff Clune for pointing HyperNEAT out to us and
              > everyone on this list - Ken Stanley especially - for the fruitful
              > discussions.
              >


            • andrei.rusu
              Hi Jeff, everyone, The biases of the output layer, which produce the default motion pattern, are also evolved using pretty much the standard approach for the
              Message 6 of 6 , Jun 21, 2010
              • 0 Attachment
                Hi Jeff, everyone,

                The biases of the output layer, which produce the "default" motion pattern, are also evolved using pretty much the standard approach for the sandwich substrates in the C++ HyperNEAT (2.6) examples, i.e. another output node for the CPPN, which means the full differentiation argument applies for the biases as well.

                Producing a default motion pattern in one direction was indeed straight-forward thanks to HyperNEAT; the actual direction of motion appeared to be established by some initial conditions, followed by a smooth gait on that track for the rest of the evaluation, in high fitness individuals.

                The real challenge was to get constructive responses from the individual modules in scenarios like: approaching a wall, or stepping over a pile of bricks. This would be hand-crafter in most real robots, or achieved via some form of reinforcement; opening evolutionary controllers to external environment inputs, and paying for coordinated, situation-specific behavior, like navigating a corridor, possibly a maze, means playing a different game altogether. However, achieving (modular) behavior this particular with neuro-evolution, via one genome, is an exciting research track for us, which we have just began to investigate.

                Best regards,
                Andrei


                --- In neat@yahoogroups.com, Evert Haasdijk <evert@...> wrote:
                >
                > Hi Jeff,
                >
                > Thanks for the interest and kind words. It was a fun experiment and we are enthusiastically moving forward with it. I'll try and get some videos out this week.
                >
                > > Your paper prompted a few questions in my mind:
                > >
                > > 1) Why did you use the 'habituation' model for inputs (where you only
                > > generate inputs values for new stimuli)? Did you only try this after the
                > > more traditional way of doing it (just feeding the raw range-finder inputs
                > > into the network) did not work?
                >
                > As you suggest, it was the result of some trial-and-error; it is fairly straightforward to achieve steady -i.e., non-reactive- locomotion without any sensors. It proved much harder to get any reactivity (i.e., negotiating obstacles) into the behaviour. Using raw range-finder inputs didn't produce that, that led us to try the habituation model. That being said, we're still working towards more prominent obstacle avoidance/negotiation.
                >
                > >
                > > 2) I really liked your idea of having a default behavior (controlled by the
                > > biases in the output layer), with the rest of the ANN only getting non-zero
                > > inputs when objects are sensed. Did you evolve the biases too, or hand code
                > > them in some way? If you evolved them, what method did you use (what were
                > > the coordinates of the source/from node fed into the CPPN when querying for
                > > a bias value)?
                >
                > All the substrate weights were determined by the evolved CPPN. I'm not entirely sure about the technical detail of the biases, I hope Andrei (who's on this mailing list as well) will answer this one in more detail.
                >
                > >
                > > 3) You say that the fact that some of the lines are different in figure 9
                > > indicates that the ANN modules are different (in their wiring and function).
                > > But wouldn't you see different lines even with identical controllers in each
                > > module because different modules are getting different inputs (e.g. when one
                > > leg moves close to a wall)?
                >
                > The spikes in the graphs are caused by obstacles coming into or disappearing from sensor range. The base levels of the graphs (particularly different in 9c) differ only because of modular differentiation. I think, but can't be certain without further analysis, that the direction and magnitude of the spikes differs in part because of modular differentiation, as well.
                >
                >
                > >
                > > 4) Did you see any symmetry (e.g. left-right) or other form of regularity in
                > > the gaits (like I see in my HyperNEAT-evolved gaits)?
                >
                > Yes we did: typically, after a certain amount of evolution, we would see symmetrical movement between the left and right legs. Not at all something that we pressed for; it just emerged.
                >
                > Cheers,
                >
                > Evert
                >
                > >
                > > Overall I loved the work and look forward to more. Thanks again for the kind
                > > acknowledgement.
                > >
                > > Best regards,
                > > Jeff Clune
                > >
                > > Digital Evolution Lab, Michigan State University
                > > jclune@...
                > > www.msu.edu/~jclune
                > >
                > > > From: evertwh2004 <evert@...>
                > > > Reply-To: "neat@yahoogroups.com" <neat@yahoogroups.com>
                > > > Date: Fri, 04 Jun 2010 15:00:34 -0000
                > > > To: "neat@yahoogroups.com" <neat@yahoogroups.com>
                > > > Subject: [neat] New paper „HyperNEAT for Locomotion Control in Modular Robots‰
                > > >
                > > > We are very pleased to announce our new publication "HyperNEAT for Locomotion
                > > > Control in Modular Robots," which will appear in the proceedings of the 9th
                > > > International Conference on Evolvable Systems (ICES 2010). The manuscript is
                > > > available here:
                > > >
                > > > http://www.few.vu.nl/~ehaasdi/papers/Modular-Locomotion.pdf
                > > >
                > > > In this paper we introduce the idea of modular differentiation with HyperNEAT:
                > > > we use a HyperNEAT encoding to generate homogeneous (in the literal sense) but
                > > > varying neural network controllers for individual robot modules that are
                > > > linked together into a larger `organism.' These controllers allow the
                > > > multi-robot to locomote purposefully and even somewhat reactively.
                > > > This paper constitutes a proof of concept and we hope to expand on this work
                > > > to allow HyperNEAT-based control in arbitrary and ultimately developing
                > > > multi-robot organisms.
                > > >
                > > > We would like to thank Jeff Clune for pointing HyperNEAT out to us and
                > > > everyone on this list - Ken Stanley especially - for the fruitful
                > > > discussions.
                > > >
                > >
                > >
                >
              Your message has been successfully submitted and would be delivered to recipients shortly.