Nvidia Jetson Nano

396 views
Skip to first unread message

Daren Schwenke

unread,
Mar 19, 2019, 10:17:49 PM3/19/19
to OpenPnP
I just got a little bit excited as a powerful machine vision platform in a small and relatively inexpensive form factor just happened.
128 Cuda cores...
It even has a Pi camera slot onboard. 
Ok... a lot excited..


This would be absolutely perfect for running OpenCV with the P1.
The P1 was my attempt at retrofitting 3D printers for PNP, which I've stopped work on for the time being, but was built to use a Pi camera as both top and bottom vision.

Anyone got one to try out?

James Shelby

unread,
Mar 19, 2019, 10:35:55 PM3/19/19
to ope...@googlegroups.com
They just got announced.  Looking at 16 weeks before availability from the vendors.

--
You received this message because you are subscribed to the Google Groups "OpenPnP" group.
To unsubscribe from this group and stop receiving emails from it, send an email to openpnp+u...@googlegroups.com.
To post to this group, send email to ope...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/openpnp/c66daa32-e338-4e15-a71d-62fa90170488%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

Daren Schwenke

unread,
Mar 19, 2019, 10:42:40 PM3/19/19
to OpenPnP
Awwww.....
Arrow shows 16 weeks.  Sparkfun doesn't have a date.  SeedStudio shows Apr 12th!
Here's hoping...

James Shelby

unread,
Mar 19, 2019, 10:48:12 PM3/19/19
to ope...@googlegroups.com
More clarity...specifically

Availability
The NVIDIA Jetson Nano Developer Kit is available now for $99. The Jetson Nano module is $129 (in quantities of 1,000 or more) and will begin shipping in June. 

Reference direct from NVidia:

Jarosław Karwik

unread,
Mar 20, 2019, 3:22:11 AM3/20/19
to OpenPnP
I have Jetson TX2 with Orbitty shield - it performs very well for for image processing ( ~ 50% of standalone 1050TI card in our case ).
But it would be too expensive I guess

Camren Carlyle

unread,
Mar 20, 2019, 8:12:12 AM3/20/19
to OpenPnP

You can get 3rd party carrier boards for the Jetson already. If you need a smaller footprint. I don't have a link, but my buddy used one on our competition robot. Worked fine, but we didn't get to the point where we implemented it for video processing. The carrier board was the size of a raspberry pi as far as length and width is concerned. It was thicc though. The secondary side of the board was just the jetson and a giant heat sink w/fan.

Daren Schwenke

unread,
Mar 20, 2019, 8:19:31 AM3/20/19
to OpenPnP
Sounds like the nano developer board.  At $99 the whole thing is priced cheaper than just the module so I think I'm going to just impatiently wait for it to drop.  
Thank you though.  Good to hear about the performance numbers.

James Shelby

unread,
Mar 20, 2019, 5:12:32 PM3/20/19
to ope...@googlegroups.com
I didnt ever see either of the other Jetson boards drop in price but I've been waiting 3 years for this board ever since the Nvidia shield came out.  The cost for just the X1 chip is insanely over priced in small quantities as it was cheaper to just buy another nvidia shield to develop on.

This is a good move by Nvidia for the cost.


--
You received this message because you are subscribed to the Google Groups "OpenPnP" group.
To unsubscribe from this group and stop receiving emails from it, send an email to openpnp+u...@googlegroups.com.
To post to this group, send email to ope...@googlegroups.com.

Maple_Dude

unread,
Mar 20, 2019, 7:22:42 PM3/20/19
to OpenPnP
What is it like to develop on these boards? Ive looked at them in the past and I think something eluded me - though I only really have embedded level programming skills so I don't know whats what for systems like this. 

James Shelby

unread,
Mar 20, 2019, 7:27:22 PM3/20/19
to ope...@googlegroups.com
It uses the linux4tegra based OS which is basically a derivative of Ubuntu.  They are more of an application board unlike the openpnp types which are more controller based with firmware.

4G of ram on the nano is going to make things really productive.


--
You received this message because you are subscribed to the Google Groups "OpenPnP" group.
To unsubscribe from this group and stop receiving emails from it, send an email to openpnp+u...@googlegroups.com.
To post to this group, send email to ope...@googlegroups.com.

Daren Schwenke

unread,
Mar 21, 2019, 3:54:01 AM3/21/19
to OpenPnP
I personally am particularly interested in the onboard CSI connector.  That gives you a 1.5GPS connection to the camera, which lets you do full HD video with zero compression basically directly to memory.  
Combined with the fast path for machine vision you get with the Cuda cores, I think this could be a very low latency solution for simple machine vision applications like we are doing here.

Daren Schwenke

unread,
Mar 21, 2019, 6:07:51 PM3/21/19
to OpenPnP
SeedStudio updated the date, and passed the Turing test...  Estimated avaiability Date: Apr 04, 2019

Daren Schwenke

unread,
Mar 22, 2019, 8:27:33 PM3/22/19
to OpenPnP
And they updated it again.  Now May 07.  :(

--
You received this message because you are subscribed to a topic in the Google Groups "OpenPnP" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/openpnp/g2OBGJ16SX0/unsubscribe.
To unsubscribe from this group and all its topics, send an email to openpnp+u...@googlegroups.com.

To post to this group, send email to ope...@googlegroups.com.

Tim 1234

unread,
Mar 28, 2019, 8:55:35 AM3/28/19
to OpenPnP
Got one coming any day. 

Not sure how feasible it will be, but do plan on using it with openpnp using the 8mp pi camera as bottom vision. 

I'm no expert on Linux or Java( c is my language), so could be a long struggle to get up and running

Daren Schwenke

unread,
Mar 28, 2019, 10:17:47 AM3/28/19
to OpenPnP
The Pi camera v1.3 returns non-discrete resolutions which breaks openpnp-capture.  
When I implemented non-discrete resolutions in openpnp-capture, I did so by returning all the possible combinations.  That... didn't go so well.  :)  Crashes openpnp, I assume due to the massive resulting array of resolutions.

So I ditched full non-discrete resolution support and implemented it to return what I knew to be the 'tipping point' resolutions.  That is where the camera switches modes, like if you set the resolution to exactly half of physical, it will automatically do 2x2 binning in hardware.  Or if you set to a 16:9 resolution it will crop the top/bottom instead of scaling.

Of course this is subjective and could change with each camera.  So I classified that as a hack and I never pushed it up.  
You are welcome to it though...

I think the simple solution is when non-discrete resolution support is detected, we should just create a box where you can set one.  That was a rabbit hole I didn't want to enter at the time.


--
You received this message because you are subscribed to a topic in the Google Groups "OpenPnP" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/openpnp/g2OBGJ16SX0/unsubscribe.
To unsubscribe from this group and all its topics, send an email to openpnp+u...@googlegroups.com.
To post to this group, send email to ope...@googlegroups.com.

Tim

unread,
Mar 28, 2019, 10:38:16 AM3/28/19
to OpenPnP
Thanks for the input.

Nano has arrived, only booted up Ubuntu so far.

Message has been deleted

Daren Schwenke

unread,
Mar 28, 2019, 7:02:10 PM3/28/19
to OpenPnP
I'm not sure I understand what you mean, but I think we mean the same thing: Keeping the data stream and image pipeline totally withing the Nano.
The Pi camera images go over the CSI connector, which provides uncompressed full rate HD video to memory, which is read by OpenCV without translation, which then uses the whole point of using the Nano, the Nvidia Cuda cores, for hardware accelerated machine vision.  
Then the actual motion commands being fulfilled by whatever...  In my case I plan on using Machinekit.
Yes,

On Thu, Mar 28, 2019 at 4:47 PM fish <lefish...@gmail.com> wrote:
Would it make sense to use the board for the upvision cv-stuff with the local mipi camera and send only vectors to the openpnp host?
I know, there would be needed a completly different driver for jetson-communication and big brake in openpnp structure, but could it be faster?
Especially when several upvision jetson boards would work in parallel - it could be a performance enhancement.


On Thursday, March 28, 2019 at 3:17:47 PM UTC+1, Daren Schwenke wrote:
The Pi camera v1.3 returns non-discrete resolutions which breaks openpnp-capture.  
When I implemented non-discrete resolutions in openpnp-capture, I did so by returning all the possible combinations.  That... didn't go so well.  :)  Crashes openpnp, I assume due to the massive resulting array of resolutions.

So I ditched full non-discrete resolution support and implemented it to return what I knew to be the 'tipping point' resolutions.  That is where the camera switches modes, like if you set the resolution to exactly half of physical, it will automatically do 2x2 binning in hardware.  Or if you set to a 16:9 resolution it will crop the top/bottom instead of scaling.

Of course this is subjective and could change with each camera.  So I classified that as a hack and I never pushed it up.  
You are welcome to it though...

I think the simple solution is when non-discrete resolution support is detected, we should just create a box where you can set one.  That was a rabbit hole I didn't want to enter at the time.


On Thu, Mar 28, 2019 at 8:55 AM Tim 1234 <timtr...@gmail.com> wrote:
Got one coming any day. 

Not sure how feasible it will be, but do plan on using it with openpnp using the 8mp pi camera as bottom vision. 

I'm no expert on Linux or Java( c is my language), so could be a long struggle to get up and running

--
You received this message because you are subscribed to a topic in the Google Groups "OpenPnP" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/openpnp/g2OBGJ16SX0/unsubscribe.
To unsubscribe from this group and all its topics, send an email to ope...@googlegroups.com.

To post to this group, send email to ope...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/openpnp/0b74111d-9ae4-47cf-a076-8573a4e37aaf%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

--
You received this message because you are subscribed to a topic in the Google Groups "OpenPnP" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/openpnp/g2OBGJ16SX0/unsubscribe.
To unsubscribe from this group and all its topics, send an email to openpnp+u...@googlegroups.com.
To post to this group, send email to ope...@googlegroups.com.

Daren Schwenke

unread,
Mar 28, 2019, 7:05:16 PM3/28/19
to OpenPnP
Um...reading again.. perhaps not.  
That's what I was thinking anyway. Feel free to detail.  :)

Tim

unread,
Mar 29, 2019, 5:32:09 AM3/29/19
to OpenPnP
That's my plan, but using a duet 3d for the gcode work.
Message has been deleted

Daren Schwenke

unread,
Mar 29, 2019, 7:17:02 AM3/29/19
to OpenPnP
For that bit, I'm currently using Machinekit on a BBB.  Communication is over ethernet.
Machinekit has a proper trajectory planner and S-curve acceleration. Not that we will really be taking advantage of that sending one command at a time..

The plan is to go to a Pocketbeagle and build a ramps type board for that. 
Communication between the Pocketbeagle and the Jetson will then be ethernet over usb or serial.

On Fri, Mar 29, 2019 at 5:47 AM fish <lefish...@gmail.com> wrote:
Yes - keeping the pipeline on the jetson to prevent actual usb-cam bottleneck.

For motion we are all waiting for a better smoothie, but with coordinated s-shaped and not trapezoid motion at higher pulse rates.

--
You received this message because you are subscribed to a topic in the Google Groups "OpenPnP" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/openpnp/g2OBGJ16SX0/unsubscribe.
To unsubscribe from this group and all its topics, send an email to openpnp+u...@googlegroups.com.
To post to this group, send email to ope...@googlegroups.com.
Reply all
Reply to author
Forward
Message has been deleted
0 new messages