Sunday 1 July 2012

Dario Floreano -- Evolution of Adaptive Behavior in Autonomous Robots


    Abstract: A spectacular demonstration of the power of natural selection comes from experiments in the field of evolutionary robotics, where scientists have conduct- ed experimental evolution with robots. Evolutionary robotics has also been advocated as a method to automatically generate control systems that are comparatively simpler or more efficient than those engineered with other design methods because the space of solutions explored by evolution can be larger and less constrained than that explored by conventional engineering methods. In this talk I will examine key experiments that illustrate how, for example, robots whose genes are translated into simple neural networks can evolve the ability to navigate, escape predators, coadapt brains and body morphologies, and cooperate. We present mostly -- but not only -- experimental results performed in our laboratory, which satisfy the following criteria. First, the experiments were at least partly carried out with real robots, allowing us to present a video showing the behaviours of the evolved robots. Second, the robot's neural networks had a simple architecture with no synaptic plasticity, no ontogenetic development, and no detailed modelling of ion channels and spike transmission. Third, the genomes were directly mapped into the neural network (i.e., no gene-to-gene interaction, time-dependent dynamics, or ontogenetic plasticity). By limiting our analysis to these studies we are able to highlight the strength of the process of Darwinian selection in comparable simple systems exposed to different environmental conditions.

    D. Floreano and L. Keller (2010) Evolution of Adaptive Behaviour in Robots by Means of Darwinian Selection, in PLOS Biology, vol. 8, num. 1, p. e100029.
    Harvey I, Di Paolo E, Wood R, Quinn M, Tuci E (2005) Evolutionary robotics: a new scientific tool for studying cognition. Artif Life 11: 79-98.

Comments invited

21 comments:

  1. This talk was really interesting. I'm not an AI professional, but I share with the author the intuition that robot must show behavioral plasticity in order to look (become?) conscious. The next step in this direction is maybe to give the robot the ability to code its own behavioral responses to achieve a goal. If we give him a goal to care about and the ability to reach this goal in creative ways, this could be a real progress.

    ReplyDelete
  2. Evolutionary robotics looks awesome! If a robot did evolve a form of consciousness (supposing we agree that consciousness is something in addition to sensing the environment), how would we know? We cannot use this technique to find the function of consciousness unless we have a technique for identifying consciousness.

    ReplyDelete
  3. I don't understand why building physical robots is necessary for this technique. Couldn't all of the physical aspects be modelled in a computer program?

    ReplyDelete
    Replies
    1. But watching physical robots is more entertaining! :) I suppose there must have been a good reason since using simulations would have gone much faster! Maybe it has something to do with Dr. Floreano's very first slide with the artistic rendition of what he wants to build someday, real flying robots lined up in scary-looking formations ready to attack!

      Delete
    2. Thanks to the new set of physics-based simulations, these days it is possible to carry out realistic experiments with robots in simulation. However, there are still important aspects where real and simulated robots are different and these differences can be picked up and amplified by artificial evolution. As a result, controllers evolved in simulations do not work in reality. Furthermore, if you do not build a physical robot to start with and work only in simulation, you may be led to simplify your simulated robot and introduce assumptions about perception and action that do not hold in the real world. Finally, even physics-based simulators cannot handle well multiple physical interactions that occur when you have multi-robot systems: this often results in the software halting unexpectedly or in computation running slower than real time.
      What I often do these days is to start by building a physical robot that fits my research goals and then simulate it with a physics-based software. Most of the evolutionary experiments are run in the physics-based simulator and the resulting controllers are tested on the physical robots to check for inconsistencies and possibly refine the simulator parameters.

      Delete
  4. I think what Floreano is doing with his group is impressive. Although I can't exactly tell if it was just for entertainment's sake, I am a big fan of the anthropomorphization of evolved robot traits: 'vanity, selfishness, sun-worship, altruism'.
    When put in the context of an experimental framework for understanding consciousness, however, I'm not sure if using these attributions will really get us closer to understanding how these traits function in humans.

    ReplyDelete
  5. I don't think the model he's using can explain how these traits function, but it does shed some light on certain behaviors that we're considered maladaptive (i.e. altruism).
    Also, if we're looking to build a machine that passes the Turing test (in this case, T3), we don't need to understand HOW to create certain functions. We only need to give AIs the proper tools for them to develop these functions through Evolution Robotics.
    We can then analyse the Turing Test candidates and see what computations create the functions we're interested in.

    That will not solve the Hard Problem, but it is a nice way to take on the easy one.

    ReplyDelete
  6. I agree with Martha, evolutionary robotics looks awesome!
    I have to look up more information and read more papers in this field, but I find the idea of, not only autonomous robots, but adaptive behaviour in robots quite fascinating. (Maybe in some way because it makes them seem more human. )

    Also, I don’t know if this is way off or has already been answered but I was wondering if we are recreating a sort of natural selection in these autonomous robots, which leads to the evolution of adaptive behaviour, how does recombinant DNA(I guess it wouldn’t be DNA in the case of robots, but whatever their equivalent is) and random mutations fit into this ‘evolution’ of robots? Do those human equivalents exist in ‘robot evolution’ or is it a completely different process?

    Izabo Deschênes

    ReplyDelete
    Replies
    1. Yes! I was also wondering -how- or -through what mechanisms- they were evolving. Did I miss this?

      Delete
  7. How do the robots "know" what to do? I realize they are programmed, but I am unclear as to the extent of their programming. Take for example the "Active Vision" robot that was moving around in the office. It was basically a pan-tilt camera on wheels, with a proprioceptive pseudo-neural network, etc. It could "maximize fitness" by moving around the environment (within the confines of the four mini-walls set up on the office floor). Some of the initial generations of robots paid attention to things that, as Dario put it, were not relevant to the task - for example, the window and the camera. If the task is to move around, and the robot is programmed to do this, why wouldn't it? If the task is to move around and the robot is not programmed to do this, then why would it?! Either way I have trouble viewing the results (no matter how fascinating) as particularly informative. I get that exploring the environment is useful for survival, and that the roboticists then "select for" this trait, operationalized as moving around. But looking is exploring too!!!

    ReplyDelete
    Replies
    1. Relatedly, the question that arose in me when watching the pan-tilt robot in the office was: Why did it "choose" to look out the window, and what made it "want" to look at the camera? Perhaps it is merely scanning the environment in order to fulfill it's objective (did it have one?!), but then what made it focus on those particular things? Chance? Why not move on? Also, I have to say that I found it oddly moving to see the robot gazing at the sun and then equally depressing to see it in it's final evolution - functioning "perfectly" with its head down staring only at the line of the floor! Crazy anthropomorphizing, I know, but it was like it had been lobotomized or something.
      (adapted from original posting on Turing Consciousness Facebook Page July 1)

      Delete
    2. This comment has been removed by the author.

      Delete
    3. The robots, as Dr. Floreano said, are governed by artificial neural networks. There is a fixed number of artificial neurons that are connected to each other according to a fixed, pre-determined graph. (It is common to build what are called ‘feed-forward networks’, where you have a fixed number of levels and every neuron on a given level is connected to every neuron on the next level, but the neurons on the same level are not connected to each other. When the network is activated, the neurons from the first level send a signal to neurons of the second level, those from the second level send it to the next and so on...) From what I understood of his talk, everything the robots do is ultimately controlled by these networks. The only thing that changes as the robot evolves is the strength of the excitatory or inhibitory connections between neurons. There are ‘motor neurons’ that are directly linked to the various engines already contained in the robot and there are ‘input neurons’ that get their activation value from various input sensors already contained in robot. At the beginning of the artificial evolution process, the motor neurons simply fire in a random way. Then, as the strength of the connections is readjusted, the network learns to use the activation value of the input neurons to determine how the motor neurons should react in order to satisfy the conditions that the modeller is selecting for. So, in a nutshell: although the architecture of the networks is fixed, as Dr. Floreano said, there is a lot of leeway for change over time. Now, the question as to whether it is right to anthropomorphize these robots is another issue altogether. It is certainly not right to do it at the beginning of the process. But, even at the end of process, it seems to me that the neural networks are still too simple for us to ascribe intentions to the robots. In most cases, they are reacting in a very simple, mechanical way to certain stimuli. They certainly don’t have the ability to acquire knowledge about the world in general.

      Delete
  8. I find it very interesting that evolutionary robotics seems to mimic findings concerning group selection versus genetic similarity in evolutionary robotics. This new field opens fascinating possibilities for both cognitive neuroscience and biology proper. Extremely interesting possibilities.

    ReplyDelete
  9. PREVIOUSLY POSTED ON FACEBOOK :

    "As an evolutionist, I found the talks on evolutionary robotics absolutely stuning, especially that of Pr. Floreano. However, I was wondering why we would like to study consciousness through robots? it appears to me that if we are doing so, it is because we kind of wanna prove that the idea of consciousness is not limited to living beings? Pr. Floreano almost convinced me that in theory, we could be able to create a robot that have a certain form of consciousness and that this form of consciousness could be very similar to that found in humans and would be able to develop human-like abilities such language, decision making and free will. It seems to me that studying consciousness through robots is the same as trying to say that if you combine the matter with the right conditions, you'll have a form of consciousness and this form of consciousness has to be completely encoded and that it should be the same for humans. Evolutionary robotics is like demonstrating that human consciousness is no more than the result of the expression of our own encoding, mainly our DNA. This encoding has to be universal but its expression can be control by environmental stimuli. Combination of bilion of environmental different stimuli let us think that life, and especially in the case of humans, cannot be totally encoded, mainly it cannot be totally determined. However, robots is the evidence that life is completely determined and predictable. Actually, nothing is random because it simply follows the laws of physics and some abilities such as free will that we attribute our species as being uniquely human are no more different that what we found in all other living beings... in that prospective, what does it mean to be human? does it has anything to do with consciousness? My intuition tend to tell me that it doesn't..."

    ReplyDelete
  10. I thought this was a really interesting direction to explore consciousness...or at least learning. One of the special things about humans is that we have such a long childhood. Although, I think children are probably conscious quite fast, so not sure how relevant this is in the topic of consciousness. Ah well: just a thought!

    ReplyDelete
  11. I’m not a robotics specialist but this discipline seems to be very interesting and useful, even though it’s still too difficult to imagine this subject applied to a consciousness explication. The experience in which robots learn not to bump into the wall is impressive. I was wondering if robots are able to use their just learnt ability within a new (totally different) environment.

    ReplyDelete
  12. PREVIOUSLY POSTED ON FACEBOOK BY SEBASTIEN TREMBLAY

    Dr Floreano’s work on adaptive behavior in autonomous robots offers, in my opinion, the best model so far presented in this conference to bring up a distal explanation to consciousness. Although it may seem counter-intuitive to work with unconscious robots to explain why consciousness exists, I believe that the power of this approach resides in the possibility to reenact evolution under experimenter-defined selection pressures. Understanding what selection pressures could favor an adaptation as abstract as consciousness would not only illuminate us on its true adaptive function, but also provide insights as to which other living entities might have evolved a consciousness as a result of a similar selection pressure history.

    Martha Shiell
    When the audience member asked about robot evolution for complex multisensory integration, it would have been interesting to speculate on the adaptatiions that robots may have developed... I think consciousness may play a role in this.

    Inge Broer
    Absolutely! I agree that it's a very interesting approach and we might indeed get to a mechanism of evolution of consciousness. However, to me, it still begs the question:
    How would such an evolved machine communicate its consciousness to us? What might it look like? If it's true that there are many consciousnesses, as Harvey kept driving home, some machines are already conscious of relevant stimuli in their environment. At what point do we say: "Yes. Within these restrictive conditions, the machine has evolved to be conscious"? And perhaps that isn't the relevant question. Perhaps better question are: what are the mechanisms for this machine to be so capable of knowing its environment? Perhaps the same logic applies to the study of consciousness in humans. But then we'd again continue studying the "easy problem" of consciousness.

    Sébastien Tremblay
    Inge, I understand your point regarding the difficulty to assess consciousness in an evolved robots, although I believe this caveat is not specific to the approach of Dr Floreano, but rather universal (robots and humans). I think Dr Damasio brought up an important point when he answered a question from the audience regarding the alleged inability for him to assess consciousness in his "patient B". He answered that assessing consciousness state of a patient is one of the first thing he learned in clinical neurology training ! I would have no reason to doubt of the consciousness of an evolved robot if it could successfully pass soma-adapted version of Damasio's clinical neurology assessment.

    Pascal Riendeau
    I disagree. I'm far from sure consciousness has an adaptive function, and if it has one, it might only be in virtue of being a necessary emergent property of information integration over time, the likes of which we see in brains. Studying the neural correlates of consciousness is much more neutral in that regard, no?

    Vincent Duhamel
    Sébastien Tremblay "I would have no reason to doubt of the consciousness of an evolved robot if it could successfully pass soma-adapted version of Damasio's clinical neurology assessment." That's way too easy. You can program a robot to disclose its hierarchy of goals or internal states. Characters in video games do this all the time. I don't think we should say they are conscious.

    Sébastien Tremblay
    Pascal, studying the neural correlates of consciousness would at best provide a proximal explanation (how?), not a distal one (Why?). The interest that this conference has toward evolution theory implies the assumption that consciousness is an adaptation. But in a strict sense you are right; consciousness remains to be proven an adaptation by contrasting the fitness of individuals with and without the phenotype.

    Sébastien Tremblay
    Vincent, the neurological examination goes a little bit beyond GTA3-like conversations.

    ReplyDelete
  13. Xavier Dery ‏@XavierDery

    Paraphrasing Dario Floreano: "Nothing in the past makes sense, exept from the evolutionary perspective." Best opening statement! #TuringC

    12:10 PM - 1 Jul 12 via Twitter for Android

    ReplyDelete
  14. Xavier Dery ‏@XavierDery

    Use cerebral lesions to understand the evolved behavior of organisms? Of course, you say. Well Floriano et al. do it on robots too! #TuringC

    12:25 PM - 1 Jul 12 via Twitter for Android

    ReplyDelete
  15. COPIED FROM FACEBOOK :

    Pier-Éric Chamberland:

    Mr Floreano's presentation, on the robot that mistakenly focus on the sun

    I understand the robot is analyzing the sun as a possible affordance?

    What fascinated me is that it concluded that it was not relevant or useful to his "objectives"


    Sarah Etezadi :


    A similar question arose in me when watching that same robot in the office. Why did it "choose" to look out the window, and what made it "want" to look at the camera? Perhaps it is merely scanning the environment in order to fulfill it's objective, as you say, but then what made it focus on those particular things? Chance? Why not move on? Also, I have to say that I found it oddly moving to see the robot gazing at the sun and then equally depressing to see it in it's final evolution - functioning "perfectly" with its head down staring only at the line of the floor! Crazy anthropomophising, I know, but it was like it had been lobotomized or something.

    ReplyDelete