OpenMarine

All,
Good morning.

After following the VendeeGlobe for 80 days, reading about it and by coincidence looking at project called DonkeyCar (self driving Lego car) I'm asking myself the following question ...

Are we able to make our boat (sort of ) autonomous via Machine Learning ... ?

Several people look at this already like:

http://www.frontiersinai.com/ecai/ecai20.../p0653.pdf

Quote:https://t-dab.com/blog/algorithm-breakdo...i-to-sail/

ai-captain.com

With all the work done on Openplotter as a data hub, we should be able to do something like this:

1. Capture (training) data

We could use sensors to measure the position of the rudder. With tiller boats you could even just use a small (raspi?) camera.
The same for main sail and jib; mount a webcam on your mast and it will analyse the shape of the sails and the position of the boom
Plus all the rest you already pull in via Openplotter
(and maybe also tides and weather; could even be afterwards)

All this creates lot's of "labeled training data"

2. Send it to Tensorflow which will "look" at all this data and will start to create a model
(question ... would it be possible to create a "data lake" of all our submitted data and create a general model to be used on all boats?

3. Once it is ready it will then be used in Openplotter to finetune your autopilot and give some advice on sail trim.

(some side notes. I'm not a data scientist or software dev. But I do read a lot about it)

Is this topic worth a new thread?

Let me know.

Bart

It may be possible to share certain knowledge between boats but it's difficult. I would try to use the shared data from multiple boats to find a better model, but continue training on each boat.

The other possible option would be to train a model with more parameters on lots of different boats, then transfer this and train a smaller set of parameters on the actual boat so it can continue to learn.

(2021-02-11, 10:16 PM)seandepagnier Wrote: [ -> ]It may be possible to share certain knowledge between boats but it's difficult.   I would try to use the shared data from multiple boats to find a better model, but continue training on each boat.

The other possible option would be to train a model with more parameters on lots of different boats, then transfer this and train a smaller set of parameters on the actual boat so it can continue to learn.

Sean i agree with you. Would be great if the OP install base could help on creating a crowd based data lake. Is that a feasible idea?
(It is actually what Tesla is doing as well, the fleet is helping the fleet (which consist of diff cars and trucks)

Hi Sean,

I have send you a private message about a week ago, did you notice?

Have a look at this video where the describe building models not just for one sheep but for numerous ("fleet" like)

https://www.youtube.com/watch?v=bdFSj0vtCFY

Who wants to further discuss?

I'm really interested in this idea. I think a google coral board might be an ideal platform to run models trained in the cloud. https://coral.ai/docs/edgetpu/models-int...-on-device

I'd love to see a solution developed that would be aware of the weather forecast and local observations, potentially sharing local observations with nearby vessels to modify the responsiveness to the current weather situation. I have the feeling that the ability to judge sea state could be a key factor in improving performance over the basic polar data.

I'm not a machine learning expert and in my naivety it seems like it would be easier to solve than many machine learning problems that have had much success such as car autopilot and large language models.

Could the first step to this be a more advanced polar chart creator, based on observations of average real world progress in various conditions? can the current method of recording a polar chart be used as a hint for a machine learning model ?

Hi all, I have just managed to get hold of one of the elusive Coral.ai Tensorflow edge units and connect it to my Openplotter Pi 4. I am thinking that if I train it to recognise navigation buoys, Marks and lights as well as recognise yachts, boats and other floating objects than a camera mounted at the bow could keep an extra eye out for collision hazards and navigational marks. Has anyone tried this?
I see three challenges, First collecting the training data that probably just means a lot of sailing and photographing.  Second estimating / calculating distance to detected objects I am not sure of the simplest way to do this so if anyone has any ideas I would be pleased to learn of them. Third, getting the object information into a real-time map overlay. 

Thanks all


Hillzzz - Moody 346, Open Plotter, Raspberry Pi 4 with Pican-M hat

scrape the web for pics of buoys, boats, ships.
how about ais data too. get course speed "et cetera".
Charts ?

Nice! 

Here's a product that does the object recognition https://sea.ai/
Starting for the bargain price of 11k euro, plus about another grand for connectivity, up to a glorious 45k for for all bells and whistles. 

They are using NVIDIA® Jetson based platforms 

Perhaps it's possible to crowd source the training data via a signalK plugin? We could really annoy people with a training data captcha asking people to pick the squares with the floating container every log in :-) 

I have a GPU server that could be used for training runs, but I'm quite new to this stuff so I don't really know where to start. 

The coral has object recognition examples available though and the demo videos do look good enough that it would be useful.

---

Maybe the distance estimate isn't that hard, a camera autofocus type system could potentially give a fairly good estimate.

https://www.youtube.com/watch?v=o9e0wAOK7JE

Of course one solution to training data is creating synthetic data sets. 

This approach seems to be gaining support for being a cheap alternative and the results are said to be very good. If you have a look at that video it's easy to see why.