Next print is a 3H print Art4AxelShaft print. I have chosen to print this in Corlorfabb HT as a test. This part probably needs more physical strength and I was running out of NGEN.
I want to be more independent and create my own models. So I started to learn Fusion 360 a few weeks ago. The main issue I have is that I have issues with the mouse control to move the object. You have to press the scrollwheel and it is hurting my right hand.
Sepio pointed out that he uses a 3dconnection mouse. So I went after a 3dconnection SpaceMouse Pro so I can control the movement with my left hand instead. Expensive! Probably too expensive (€340), but ergonomics is more important if it affects your health.
I initially was thinking about the simple one. But when I looked at the reviews it comes off from the table. (I also wants to use it for games) You also did not have a resting pad.
I learned that there is some learning curve, but I don't think I will have an issue with this because my left hand has always been better at steering.
Experiments with ColorFabb XT, HT, NGEN, Digital Rise PETG, and ABS.
I think for mechanical parts NGEN may be the best choice to be used used in the Thor robot. Can support 85 degrees and is pretty close to the feel as PLA but more strong and less bendable.
In the Prusa MK2/S I can print NGEN at a speed of 70 mm/sec.
ABS (ICE) has way too many warping, and failed prints.
Just ordered more filament ColorFabb_XT, ColorFabb_HT, and NGEN from om3d. Since the Firefox upgraded to the latest version their web site works perfectly.
Holidays are coming and a fully functioning 3D printer that is waiting for jobs.
Yesterday I did a test-print my ABS (white ICE) was a disaster. But somehow I also have ABS (red ICE) that appears to print better.
PETG (Digital Rise) is good but too soft, CorlorFabb NGEN is spot on, but I must print more top layers (4 total). NGEN did give me cuts when I manipulated it because it has sharper edges than ABS or PETG. Closer to PLA.
So far all my ColorFabb NGEN printed without issues once calibrated.
The price is what expensive. But then again so far every print succeeded.
The Carbon fiber, I am going to wait with that one. Not very healthy for my printer. I had loops sticking out and it misaligned the head when it struck the carbon filament. I have to ask my colleague how he does it.
One thing I did notice with thee white NGEN. I seem to have dark islands. I think i is dust buildup that somehow gets stuck in the extruder and then suddenly releases. It could also be fragments from the last carbon fiber filament. I never noticed this with white ABS printing.
Still busy with the Movidius Neural Stick and the Raspberry Pi. The hardest part of learning TensorFlow not the theory but the misery of installing the software to run it. Somehow after my forced Windows update stuff is broken again. In the Raspberry Pi it seems to work good.
Very productive 3 days off. I have set up a Raspberry Pi installing all software for TensorFlow and the Movidius Neural stick. From now on I can create my own neural networks on that device. And also printed my Raspberry Pi case. :-)
Upgraded my Original Prusa MK2 to the MK2/S upgrade upgrading the mechanical properties. Better ball bearings and better rods. But I had to disassemble and reassemble a big part of the printer.
Now we can go back and focus on Thor itself in the next weeks to come. (Still more holidays to take)
Movidius Neural stick + Raspberry camera requires a 32 GB Sd card.
After the raspberry pie Linux is installed things gets interesting to make the stick function.
* Install the SDK: Quick
* Make Install (the Movidius SDK) : 1 hour
* Make examples (the Movidius SDK examples) : 6 hours
The raspberry camera seems to need a firmware update to get it working.
The Movidius™ Neural Compute Stick can now use Tenserflow Inception V3 (Google trained base) and if you feed it an image then it can recognize what is on that image.
This is only the starting point, we can now add your own images and your own categories.
What I am going to do with this stick I do not know yet, but it could become interesting as a brain for the Thor robot. The Arduino are the muscles. The Raspberry pi the brains.
I finally managed to get the Movidius™ Neural Compute Stick functioning on a Raspberry pi. However a 16 GB SD card was too small and too slow. So I now am reinstalling the software with a 32 GB SD faster card.
It is really tough to get all the software modules installed to get TensorFlow actually functioning on Linux! But once set up everything seems to be smooth.
This AI is actually not that hard to understand. You don't need to understand all this maths, just copy and past the examples and off you go. The fight I am having is to make the software install and function.
The Movidius™ Neural Compute Stick is giving me head aches to get it functioning on an fresh new installed Ubuntu 16.04 (using VMWARE). It recognizes the stick but when I run the program it loses connection and then I get an error message. as far as I have understood, it starts up with USB 2.0 mode and when it has to run it switches to USB 3.0 for fast speed.
But to be honest it is designed to run on a Raspberry pie 3. On PC's you have more CPU powered, this stick is intend to speed up on devices that operates on low power or battery.
So I ordered my Raspberry Pie and a camera.
The cool thing I learned is that you can install the Google TensorFlow on your Windows 10 PC.
Google TensorFlow is the SDK from Google to create neural networks. It is basically equivalent to ASP.NET where you create C# code to generate html code. With TensorFlow Python/C++ is used and the output is called a "Graph". This Graph can then be compiled for the Movidius stick.
This Graph is not one giant pile of neurons (but grouped neurons in modules that you could program yourself. It is the equivalent of creating methods in conventional programming. But in this case these methods are intended to be trained for one specific job.
One of the cool things about Google is that it also created Inception-V3. This Inception is pre-trained AI that can recognize objects in images/video. So you start with this model and train it from here for your own purpose. https://www.tensorflow.org/tutorials/image_recognition
This installs Python, every single Linux package needed by Python to execute scientific code. Once Anaconda navigator is started you can choose packages to install like TensorFlow.
Something also very cool to install in Anaconda is Jupyter. http://jupyter.org/
Jupyter can create notebooks that looks very similar like Mathematics. You can now easily create Python code and test them. e.g. TensorFlow Python code code.
I am still in this learning phase of AI but I think this will become useful for the Thor project. Object image recognition is one example. But you need to add an Raspberry Pie in this case.
The Movidius™ Neural Compute Stick bay be in tomorrow. And since I have a few days off that would be perfect. This stick can be used in small projects and have off-line AI neural network capabilities.
I am wondering if we can use it to help the Thor control capability to compensate for movement.
This is technically a normal math processor but optimized for AI algorithms. They call it a VPU, very similar like a GPU and is mainly designed for visual image recognition.
This stick can now be added to a small Raspberry pie.
This Fusion 360, I like but there are clearly bugs in it that drove me nuts yesterday. You click on something and the back rectangle you just drew jumps to the other side of the axis. There was nothing that I did that would do that.
I don'l like the scroll wheel button to rotate. It keeps on zooming in every time I select that button.
Next test ABS, with object cooling fans and Thor part in ABS that always failed with a thermal runaway or warping. I am wondering if the copper alloy heater block has it now under control.
Olaf Baeyens
Discussions
Become a Hackaday.io Member
Create an account to leave a comment. Already have an account? Log In.
Sepio's Art4Lowerbody_01 has also been printed in NGEN. Perfect print and took 5H45.
I need more NGEN, but the budget constraints me to order it only from January on. The same for some parts that are being used in Sepios version.
Are you sure? yes | no
Just printed the Art3UpperBody NGEN. An 8h30 job. It has a very solid feel and no warping.
I like the air venting holes.
https://hackaday.io/project/26341-building-and-improving-the-thor-robot-arm/log/70087-day-18-redesign-of-art3-and-art4
Next print is a 3H print Art4AxelShaft print. I have chosen to print this in Corlorfabb HT as a test. This part probably needs more physical strength and I was running out of NGEN.
Are you sure? yes | no
I want to be more independent and create my own models. So I started to learn Fusion 360 a few weeks ago. The main issue I have is that I have issues with the mouse control to move the object. You have to press the scrollwheel and it is hurting my right hand.
Sepio pointed out that he uses a 3dconnection mouse. So I went after a 3dconnection SpaceMouse Pro so I can control the movement with my left hand instead. Expensive! Probably too expensive (€340), but ergonomics is more important if it affects your health.
Are you sure? yes | no
I use the most simple one. https://www.3dconnexion.eu/products/spacemouse/spacenavigator.html
It’s on the left side of my keyboard. My mouse is on the right side.
Are you sure? yes | no
I initially was thinking about the simple one. But when I looked at the reviews it comes off from the table. (I also wants to use it for games) You also did not have a resting pad.
I learned that there is some learning curve, but I don't think I will have an issue with this because my left hand has always been better at steering.
Are you sure? yes | no
Screws, lots of screws. Just ordered 1600 of them in every possible length (M3)
Are you sure? yes | no
I decided to start printing Sepio's upper part of the Thor robot. But in NGEN.
First part is a 7 hour job Art3LowerBody.factory.
https://hackaday.io/project/26341-building-and-improving-the-thor-robot-arm/log/70087-day-18-redesign-of-art3-and-art4
Are you sure? yes | no
Experiments with ColorFabb XT, HT, NGEN, Digital Rise PETG, and ABS.
I think for mechanical parts NGEN may be the best choice to be used used in the Thor robot. Can support 85 degrees and is pretty close to the feel as PLA but more strong and less bendable.
In the Prusa MK2/S I can print NGEN at a speed of 70 mm/sec.
ABS (ICE) has way too many warping, and failed prints.
Are you sure? yes | no
Just ordered more filament ColorFabb_XT, ColorFabb_HT, and NGEN from om3d. Since the Firefox upgraded to the latest version their web site works perfectly.
Holidays are coming and a fully functioning 3D printer that is waiting for jobs.
Are you sure? yes | no
Fantastic, would be nice the get some feedback of the XT and HT filaments :-)
Are you sure? yes | no
Yesterday I did a test-print my ABS (white ICE) was a disaster. But somehow I also have ABS (red ICE) that appears to print better.
PETG (Digital Rise) is good but too soft, CorlorFabb NGEN is spot on, but I must print more top layers (4 total). NGEN did give me cuts when I manipulated it because it has sharper edges than ABS or PETG. Closer to PLA.
So far all my ColorFabb NGEN printed without issues once calibrated.
The price is what expensive. But then again so far every print succeeded.
The Carbon fiber, I am going to wait with that one. Not very healthy for my printer. I had loops sticking out and it misaligned the head when it struck the carbon filament. I have to ask my colleague how he does it.
One thing I did notice with thee white NGEN. I seem to have dark islands. I think i is dust buildup that somehow gets stuck in the extruder and then suddenly releases. It could also be fragments from the last carbon fiber filament. I never noticed this with white ABS printing.
Are you sure? yes | no
Still busy with the Movidius Neural Stick and the Raspberry Pi. The hardest part of learning TensorFlow not the theory but the misery of installing the software to run it. Somehow after my forced Windows update stuff is broken again. In the Raspberry Pi it seems to work good.
Are you sure? yes | no
Very productive 3 days off. I have set up a Raspberry Pi installing all software for TensorFlow and the Movidius Neural stick. From now on I can create my own neural networks on that device. And also printed my Raspberry Pi case. :-)
Upgraded my Original Prusa MK2 to the MK2/S upgrade upgrading the mechanical properties. Better ball bearings and better rods. But I had to disassemble and reassemble a big part of the printer.
Now we can go back and focus on Thor itself in the next weeks to come. (Still more holidays to take)
Are you sure? yes | no
For the record!
Movidius Neural stick + Raspberry camera requires a 32 GB Sd card.
After the raspberry pie Linux is installed things gets interesting to make the stick function.
* Install the SDK: Quick
* Make Install (the Movidius SDK) : 1 hour
* Make examples (the Movidius SDK examples) : 6 hours
The raspberry camera seems to need a firmware update to get it working.
The Movidius™ Neural Compute Stick can now use Tenserflow Inception V3 (Google trained base) and if you feed it an image then it can recognize what is on that image.
This is only the starting point, we can now add your own images and your own categories.
What I am going to do with this stick I do not know yet, but it could become interesting as a brain for the Thor robot. The Arduino are the muscles. The Raspberry pi the brains.
Are you sure? yes | no
I finally managed to get the Movidius™ Neural Compute Stick functioning on a Raspberry pi. However a 16 GB SD card was too small and too slow. So I now am reinstalling the software with a 32 GB SD faster card.
It is really tough to get all the software modules installed to get TensorFlow actually functioning on Linux! But once set up everything seems to be smooth.
Are you sure? yes | no
This AI is actually not that hard to understand. You don't need to understand all this maths, just copy and past the examples and off you go. The fight I am having is to make the software install and function.
Are you sure? yes | no
The Movidius™ Neural Compute Stick is giving me head aches to get it functioning on an fresh new installed Ubuntu 16.04 (using VMWARE). It recognizes the stick but when I run the program it loses connection and then I get an error message. as far as I have understood, it starts up with USB 2.0 mode and when it has to run it switches to USB 3.0 for fast speed.
But to be honest it is designed to run on a Raspberry pie 3. On PC's you have more CPU powered, this stick is intend to speed up on devices that operates on low power or battery.
So I ordered my Raspberry Pie and a camera.
The cool thing I learned is that you can install the Google TensorFlow on your Windows 10 PC.
Google TensorFlow is the SDK from Google to create neural networks. It is basically equivalent to ASP.NET where you create C# code to generate html code. With TensorFlow Python/C++ is used and the output is called a "Graph". This Graph can then be compiled for the Movidius stick.
This Graph is not one giant pile of neurons (but grouped neurons in modules that you could program yourself. It is the equivalent of creating methods in conventional programming. But in this case these methods are intended to be trained for one specific job.
One of the cool things about Google is that it also created Inception-V3. This Inception is pre-trained AI that can recognize objects in images/video. So you start with this model and train it from here for your own purpose. https://www.tensorflow.org/tutorials/image_recognition
In order to make TensorFlow install on your Windows 10 or Ubuntu is by Installing Anaconda: https://www.anaconda.com/distribution/
This installs Python, every single Linux package needed by Python to execute scientific code. Once Anaconda navigator is started you can choose packages to install like TensorFlow.
Something also very cool to install in Anaconda is Jupyter. http://jupyter.org/
Jupyter can create notebooks that looks very similar like Mathematics. You can now easily create Python code and test them. e.g. TensorFlow Python code code.
I am still in this learning phase of AI but I think this will become useful for the Thor project. Object image recognition is one example. But you need to add an Raspberry Pie in this case.
Are you sure? yes | no
The Movidius™ Neural Compute Stick bay be in tomorrow. And since I have a few days off that would be perfect. This stick can be used in small projects and have off-line AI neural network capabilities.
I am wondering if we can use it to help the Thor control capability to compensate for movement.
Are you sure? yes | no
Finally found a Movidius™ Neural Compute Stick that can be ordered (82 euros). Waiting for months to get one.
https://developer.movidius.com/buy
This is technically a normal math processor but optimized for AI algorithms. They call it a VPU, very similar like a GPU and is mainly designed for visual image recognition.
This stick can now be added to a small Raspberry pie.
https://movidius.github.io/ncsdk/
I have no direct usage for it but I want it to learn all about AI that is not connected to Google services.
Are you sure? yes | no
This Fusion 360, I like but there are clearly bugs in it that drove me nuts yesterday. You click on something and the back rectangle you just drew jumps to the other side of the axis. There was nothing that I did that would do that.
I don'l like the scroll wheel button to rotate. It keeps on zooming in every time I select that button.
Are you sure? yes | no
For navigation I use a 3DConnexion Spacenavigator. That works perfectly.
I don’t recognize the problem you describe.
Are you sure? yes | no
I heard from a colleague that when you work in a sketch that bugs are known.
Are you sure? yes | no
I am looking at your 3DConnexion Spacenavigator. This controller may be interesting to be used as a manual robot controller.
Mode 1: I look at the robot and use Inverse kinetics to rotate the hand in rough mode.
Mode2: I look at the object to grab and the motion is now fine grained to grab that object also using inverse kinetics.
Are you sure? yes | no
Learning Fusion 360, time to learn to build my own models.
Are you sure? yes | no
Test fail, ABS already start warping. But the printer is functioning perfectly.
Are you sure? yes | no
Next test ABS, with object cooling fans and Thor part in ABS that always failed with a thermal runaway or warping. I am wondering if the copper alloy heater block has it now under control.
Are you sure? yes | no
* V6 Gold HotEnd: CHECK
* Fusian 360 installed: CHECK
* 5 none stop of holidays: CHECK.
* Smoke test: NO SMOKE!
* Olsson ruby nozzle: CHECK
* PLA calibration CHECK
* ABS test print: BRB.... in 5 hours ;-)
Are you sure? yes | no