[TriEmbed] If you were looking at a deep learning rover kickstarter, what would you want to see?

[Charles West]

Hello!

I'm still working on my sidewalk robot up north.  Currently its doing very
well on staying on the sidewalk/handling crossings but not so much on
listening to which crossing to take to get to where I want it to go.  TBD.

I'm working on this primarily for my thesis, but talking with my MBA
brother brought up the possibility of potentially open sourcing what I've
built (after some polish) and trying to do a Kickstarter for "A Deep
Learning Robot that can intelligently navigate to a given GPS coordinate"
(GoodBot Go?, GoodBot Buddy?).  Initial apps for it would be "Allow anyone
with a phone app and my permission to summon the robot to them and then
send it to one of the other people with the app installed" or "Go where my
XBee basestation says to".  There are several startups that seem to have
robots with this sort of capability, but as far as I know none of them are
selling the actual platforms for people to build there own ideas (like the
golf course deliveries I am looking into).

My impression so far is that there is a lot of work required to get started
in deep learning.  My current workflow has me drive the robot, transfer the
data it collected (which I display on a mapbox map of the world in a GUI),
export/transfer the training data, train and then export the trained model
back.  Even with a somewhat streamlined workflow, it can be a bit
involved.  However, the practical upshot is that if I want my robot to
dodge a particular type of object (or eventually generalize), all I have to
do is record driving around it several times.

I think it would be possible for me to simplify the workflow a lot.  First
off, for the same vehicle type training data could be aggregated from a
bunch of users and open sourced (which could be integrated into the GUI or
the robot itself).  Pretrained models could be used (and periodically
updated) if all someone wants is to use the "get to GPS point
functionality" or to speed up specialized training.  It is also potentially
possible for me to integrated the training/exporting capabilities into the
GUI (C++), robot (really slow) or a cloud service for a fee so that all you
have to do is click a button and it will export, choose model and start
training.

It would also be possible to integrate additional sensors and actuators via
config files and try to do the "do what I did" pattern with them as well.

The main questions:
1. Would a price tag of $1200 for a open source vehicle approximately this
size (https://i.ytimg.com/vi/5vPJ3CN3e9c/maxresdefault.jpg) that can learn
and intelligently navigate using sidewalks be reasonable?
2.  Do you think this sort of thing would get funded?
3. What features would you really care about?  Go to point functionality,
phone app, push of a button training?

Thanks,
Charlie West
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.triembed.org/pipermail/triembed_triembed.org/attachments/20190606/0099f9e1/attachment.htm>