Beyond DCC++

Mo-Pac Oct 15, 2020

  1. SLSF Freak

    SLSF Freak Staff Member TrainBoard Supporter

    1,141
    460
    37
    Super interesting discussion! When I think of all the stuff AI is doing these days, for MRR'ing I'm thinking of a system that uses a camera trained on the whole layout from above (or maybe a system of cameras for large layouts. A consumer model RR AI package would already be trained in a generic sense on recognizing movement of a train, identifying "units" moving (so it could identify if a train splits when the unit is broken up), and it can identify the sounds of a derailed car. With high enough resolution on the camera it could also identify individual cars and locomotives. To map the layout you do an initial setup using a locomotive carrying a IR reflective flag and run it along the layout so the system records/draws every possible route. With the combination of having a map of your layout now, coupled with visual and audio feedback the system should be able to figure out everything it would need to operate without the need of complicated sensors or wiring. Inventory software can tell it which pieces of rolling stock are equipped with decoders and sound. Adding signals and stations or way points to your layout map can allow automated control of those.

    I do software for my day job (not AI though) and sometimes I try to look at how humans do things first to have the computer then programmatically replicate that action to automate a difficult task. In our brains, we have a map of our layout, we recognize when a train splits up, we can tell if it's moving fast, slow or not at all, we know sometimes we can't see a derailment but we can hear it before the whole thing topples over. So stuff like that is how I'd approach an AI solution. The tech is already out there doing similar things (self driving cars, object and facial recognition), it's just never been put together in this way. I'm truly blown away with what AI is doing these days - I wish I had the chops to program for it but I think I'm in the early stages of aging out of the field. I can still dream big though.. (y)

    Mike
     
    Mo-Pac likes this.
  2. Mo-Pac

    Mo-Pac TrainBoard Member

    450
    312
    8
    OH WOW!! Now you are talking! I like how you expanded this idea!


    Sent from my iPhone using Tapatalk
     
  3. Mo-Pac

    Mo-Pac TrainBoard Member

    450
    312
    8
    Seriously Mike only now that we need to find someone who is a programmer and can generate the ideas from you and I. To make it work. I agree that today’s model railroading is good to the point. But having all of these excess wires and sensors that would drain the wallet by needing this equipment to be connected to a sensor that’s literally from the 1970’s or eighties! To have the functions like we’re thinking of and put them into a program. The way I see it is to keep it simple. Not by adding barcodes and barcode scanners that will take up extra room on the layout. Especially whenever digital cameras with high resolution and the AI can detect the design of the tracks by just running a train with a one time IR reflective flag over every track or by entering the pieces manually with the lengths to be able to match the layout. The added audio and visual feedback system would make it very helpful to detect derailments. The inventory software added would help determine the actual length of the whole train with the AI knowing what every piece of rolling stock’s length is. Between the sensors feeding into the tracks and the camera above it will determine the precise location of every train(just in case if the camera is malfunctioning or the tracks are dirty the AI can use one of the other if an issue has arisen it would also notify the operator that the tracks will need cleaning and the camera needs to be put back online)and know where everything is located on a layout. We just need to find someone who can put this level of AI together on software. Too me the less gadgets and wires sucking up electricity all spread out over the layout, the less costly model railroading would be.


    Sent from my iPhone using Tapatalk
     
  4. BigJake

    BigJake TrainBoard Member

    154
    94
    8
    I think too many people wave the magical "AI" flag and declare a solution is at hand, and that it's just a matter of programming...

    I have already outlined a means, which as Mo-PAC has declared, only needs to know the location of the locomotive(s) and the caboose/FRED, using specifically spaced, IR reflective (and visually transparent) dots on the track bed, that are sensed with IR emitter/detectors embedded in the locos and FREDs/cabeese, enables on-board railcom-compatible decoders, with minimal embedded firmware, can determine where it is on the track, and then report back said location to a railcom compatible command station.

    While the two quadrature strips of IR markings can be put down with a simple paint wheel, the third strip of markings (providing an absolute position reference via a pseudo-random sequence of marks and spaces) would be more difficult to accomplish. Perhaps a specially designed railcar could create that strip after the two quadrature strips are already laid down.
     
  5. Mo-Pac

    Mo-Pac TrainBoard Member

    450
    312
    8
    How much did you spend on all of these gadgets and extra wires and electronics? Yes, the items that you have are readily available since the eighties. We are in a different century. Technology has changed. Though people haven’t. This is what brainstorming is about. Putting ideas that are not thought of yet and then making it work.


    Sent from my iPhone using Tapatalk
     
  6. BigJake

    BigJake TrainBoard Member

    154
    94
    8
    I haven't spent a dime more on hardware than you have on an AI solution. It is just a proposal.

    AI is only SW, which also requires hardware (and other SW) to train, to execute, and for sensors (whether monitoring current or time-of-flight of signals) to give the AI the data it will process.

    AI is not magically better than other technologies on every problem. It does solve a lot of problems that do not yet have viable, more conventional solutions. A likely characteristic of problems for which AI provides a viable solution, is that they involve a LOT of data. Data for training, and data operate on once the algorithm is learned.

    If you really have a burning desire to apply your 21st century AI hammer to this problem, I suggest you start with processing video, from cameras overhead, and in tunnels or other features that would hide the train from cameras above.

    A lot of folks treat AI like a hammer, but not every problem is like a nail. Hammers have been around for thousands of years, and they are still better at driving nails than AI, and will likely continue to be for a long time.

    I wish you luck...
     
    Mo-Pac and Doug A. like this.
  7. Mo-Pac

    Mo-Pac TrainBoard Member

    450
    312
    8
    First of all thank you! You are absolutely correct about people using AI as an hammer. The way I see it being used as an nail instead of the hammer. Making it the whole thing as a true AI. Will take a lot of programming. The only way that AI can be magical if the program was written to self learn from task, mistakes, reactions and the different scenarios that can be possibly run. Oh yes! The AI will depend upon how good the programmer themselves is also.


    Sent from my iPhone using Tapatalk
     

Share This Page