5 ms·
Why all the robots follow the same path even though is not the optimal one? Why no robot is trying anything new? Where is the exploration/exploitation side of t
by vicpara 12y ago
Why all the robots follow the same path even though is not the optimal one? Why no robot is trying anything new? Where is the exploration/exploitation side of the strategy? Where is the additional information each robots is contributing to? Why aren't there groups of agents taking over a part of subproblem ?
If they don't use any centralized decision making then it should have taken them much less time.
The way the robots arrange in shape depicts the way a human would do it. A realistic, real time, no centralized decision making algorithm would move all the robots at the same time. Yes, the stabilization time, until all robots find their rightful position based on peer-peer communication, would take a while and that would have been acceptable. This thing over here looks a lot a lot like hard coded, human minded approaches to a trivial problem.
If the real advancement was the robot mechanical movement then they should have said so. The algorithm they have could have been done by any CS background high school student.
- scott_s 12y agoYou are relying on the popular-press description of the algorithm to assess its sophistication. If you would like to read it, I can send you the supplementary pdf which explains the algorithm in detail. The algorithms do look straightforward compared to machine learning approaches, but that's a good thing! I'm guessing you're a machine learning person? I believe this approach comes more from the control theory side of automation. Simple-looking algorithms that achieve complex phenomenon is interesting. The authors also provide proofs that the algorithms will, in fact, achieve the goal. I also want to point out that they are under hardware constraints which preclude approaches which require large amounts of computation and memory. From the supplementary material: The Kilobot robot has strict limits on its available memory: 2K RAM and 32K for program memory, which includes the code for the bootloader/wireless programming, robot move- ment, and communication libraries. The full self-assembly algorithm takes approximately 27K of program memory including 241 bytes for the shape description and scale factor. The Kilobot also has limits on message size; we optimize by combining messages from the various primitives, placing all the data in a single 7 byte message. In general, I find it dangerous to make conclusions of the novelty of research from the popular press description of it.
- angersock 12y agoMind shooting me that PDF? I had a colleague that worked with the kilobots last year while we were preparing our own paper. :)
- Retric 12y agoI don't want to downplay what was involved as I think it's a cool project. However, from a real world perspective: If the algorithm requires everything to be in one clump it's not robust even if it's 'provably' correct. It requires specifying a small set of special bots which is external control. It scales while a bird flock can respond quickly even at 100,000 birds and a simple algorithm try and calculate how long building a shape would take with 100,000 bots.
- scott_s 12y agoAll good points, but you are explaining why this research is not a landmark development. Most research, however, is iterative. Get a simple step working, which contains a bunch of simplifying assumptions. Then slowly try to remove those assumptions. That takes time and effort, and we publish each of those steps, and call them "novel".