2024/12/20: early idea to add tool changer as well
2024/12/16: adding more details with multiple extruders on same X gantry
2024/12/06: early ideas about single layer segmenting
2024/12/05: starting write-up
Introduction
While working on theoretical side of parallel 3D printing, I used the example of multi-gantry or Multi Gantry (MG) setup where a single bed is shared in Y – here a few rough sketches:
Multi Gantry with same order, Common Y Rails (CYR)Multi Gantry with same order (closeup), Common Y Rails (CYR)Multi Gantry with ability to change order, Dedicated Y Rails (DYR)Multi Gantry with ability to change order (closeup), Dedicated Y Rails (DYR)
Gantry Details
There are several challenges, such as the Y- and the XZ motion system and making the thickness of each gantry as thin as possible to have as little “dead” or “unusable” space as it adds up and also limits the printable XY space.
Y motion system
We can share a single rail where each gantry rides hence Common Y Rails (CYR), then they can’t change the order, in that case the belt routing becomes the main issue of concern – whereas when each gantry has its own rail hence Dedicated Y Rails (DYR), and each gantry has a distinct height they can change the order (lift Z to the top and move over the lower one – hence each gantry has a different Z height, but with the filament and cabling the entanglement needs to be closely observed).
Single Rail, Multiple Gantries aka CYRDedicated Rails per Gantry aka DYR
XZ motion system
We require to have two positioning axes to integrate here, the X and Z:
two independent motion systems (X axis = 1 motor, Z axis = 1 motor)
Core XZ system also using two motors, but having XZ motion combined
Jon Schone’s (ProperPrinting) 2 Gantry System: X & Z axes with each their own motor, sharing a single Y rail
Slicing for Multi Gantry System
As I realized earlier, it makes no sense to statically segment Y-space for each gantry (e.g. Y-size/n) although it would make coding easier, but seamless printing would be impossible – in reality, we need slightly overlapping spaces to achieve zero space seams where the gantries print for the same piece, so we naturally choose flexible non-overlapping operations.
Layer Segmenting
Obviously one of the simple optimization is when we are looking at a single layer, and segment it area-wise to n-amount of gantries. For sake of consistency, a wall is assigned to a single gantry to have a seamless wall/perimeter.
Y Segmenting Single Layer for 4 gantries & nozzles (red, green, blue & yellow)
Infill areas are Y segmented
each area (e.g. A1-A4) is the same to distribute work equally
Walls/perimeters are assigned to a single gantry
if walls intersect each other in Y, those are printed sequentially not in parallel to avoid Y collision
The layer segmenting approach makes it rather easy so all nozzles have the same Z, thereby segmenting print jobs becomes easy as well.
Mixing Multi Gantry (MG) with IMEX (IDEX & ITEX etc)
We can also mix Multi Gantry (MG) with IDEX (Dual) or ITEX (Triple), or in general Independent Multi Extruders (IMEX), something like this:
Multi Gantry with IDEX: Independent Dual Extruders per single XZ gantry
Layer Segmenting of Multi Gantry (4: red, green, blue, yellow) with IDEX (2) per Gantry
And like-wise segmenting layer areas in Y as before, and additionally segment in X for each printhead on Independent Multi Extruders (IMEX), in this example it’s IDEX (Dual).
As IMEX looks like scaling print parallelism further, one has to be aware of the spatial overhead for each printhead, e.g. a single printhead occupy ~40mm width in X, and similar in Y.
MG-MIEX 3×2
Early draft with more details based on Ashtar M, extended to MG, with Common Y Rails (CYR), with IDEX:
MG-MIEX 3×2IDEX Detail
Challenges
1) When doing a MG-MIEX 3×2 (3 gantry with 2 extruders each):
X: 3*2 motors = 6 motors
Y: 3 motors
Z: 3 motors
E: 3*2 motors = 6 motors we end up 6+3+3+6 = 18 motors to coordinate & control.
2) Slim design of the gantry to reduce dead-space in Y.
3) Slim printhead design to reduce dead-space in X.
4) Slicer requires to coordinate 6 printheads (e.g. T0-T5) in non-colliding and efficient way.
Solutions
Slicer
A dedicated slicer is required which segments each layer into printable areas A1 to An whereas n is the amount of gantries, and those areas might be further segmented by m-amount of printheads on the same gantry, so they share the same Y position but can have distinct X position but not switch order or collide.
As proposed above, a single printhead of a gantry is then assigned to print within the same Y-segment the walls/perimeters to have a clean wall, hence, the infills are distributed among all the printheads.
Firmware & Controller
The controller with its firmware we have a few options:
M400: wait for current moves to finish (when using T<n> with G1s together, applies to X, Y, Z, and also E)
M596 P<n> selects motion queue, prior each G0 or G1
M597: collision avoidance (at firmware level)
M598: sync up multiple motions (faster waits for the slower)
Klipper
Marlin
one controller per gantry, orchestrating between controller needed
simple setup, reliable
As the operation of the gantries are overlapping, they cannot have relative position, but rather have absolute Y position (relative position used to detect collision), the same for the X position.
Adding Tool Changer (TC)
In order to support a tool changer, the printheads need to reach a common position in order to deposit and pick up tools. In case of a single Y rail setup this seems at first sight challenging, only a dedicated tool change per gantry, at a particular a YZ position for example. The multi Y-rail setup allows any gantry to reach a common region in YZ, and so any printhead could deposit and pickup a tool from the common toolset.
2024/12/04: completing first table on printheads & nozzles
2024/07/30: starting write up
Introduction
When depositing material through a nozzle, the variables to compose a workpiece depends on the amount of nozzles and their operational spaces – let’s lay out the different methods which gives us the foundation to tackle then parallel procedures in the next part in the series.
Printheads, Nozzles & Operational Space
Printheads
Nozzles per Printhead
Nozzle Size [mm]
Layer Height [mm]
Material
Operational Space
Single nozzle FDM/FFF
1
1
0.1-1.0
0.1-0.6
Polymer (PLA, PETG, ABS)
100%
Dual nozzle FDM/FFF aka IDEX
2
1
0.1-1.0
0.1-0.6
Polymer (PLA, PETG, ABS)
2x 50%; horizontally separated
Duplex F2
2
1
0.1-1.0
0.1-0.6
Polymer (PLA, PETG, ABS)
2x 50%; vertical separated
CM3P Dual Conical
2
1
0.1-1.0
0.1-0.6
Polymer (PLA, PETG, ABS)
2x 50% of negative cone
Resin SLA (UV Laser)
1
1
0.150
0.050-0.150
Resins
100%
Resin MSLA (UV & LCD)
1
50M-100M
0.020-0.040
0.050-0.150
Resins
100%
Quantica NovoJet
1
96
0.050
Resins
100%
Stratasys J55 PolyJet
1
192*)
0.2*)
0.18
Resins
100%
Selective Laser Sintering/Melting
1+
1
0.1
0.05-0.10
Polymer or Metal Powder
100%
Stratasys J55: nozzles & nozzle size based on J850 specs, J55 details specs seem not publicized (2024/12)
Print Base
A print base is where the nozzle can extrude on. For the first layer, there is the print bed, after the first layer the workpiece or support structure can be build upon. One can alternatively use a stabilizing medium like silicon and extrude in such liquid medium which operates as bed or foundation like Rapid Liquid Printing (RLP) does:
Extruding into supporting materialExtruding into supporting material
The extruded material just has to stay where it was put, either a solid bed or a medium which prevents it to float out of position, or as traditionally printed on a print bed or base, very similar does Xolography where the solidified resin stays put as well.
Massive Parallel Nozzles Printhead
Resin printing with a printhead may have hundreds or even thousands of nozzles, yet, they share the same operational space, but due the parallel setup the print speed multiplies direct with the amount of parallel nozzles on the same printhead.
As mentioned above, we can also view Masked Stereolithography (MSLA) resin printing as a massive parallel nozzle setup, where each pixel is either an active or inactive nozzle depositing a voxel.
Anycubic Mono M7 Max MSLA Setup
Separated vs Shared Operational Space
Disclosure: I have been contracted to work on the Duplex F2 software stack (2022-2024).
Let’s take a look at the Duplex F2 printer where space is separated vertically, or the CM3P dual conical printer where the cone space is separated horizontally, or the Multi Gantry 3D printer by Proper Printing (Jon Schone). We have two printheads which never collide due their separated operational space, the firmware is simple and path planning is simple, both heads pretty much can operate independently.
When using more than two printheads it is beneficial to share the operational space, yet assume 6 or 8 printheads, each printhead needs rods to keep the printhead and position and orient the nozzle(s), so overlapping operational space requires extreme well planned tool paths avoiding any collision of the printheads.
Regular Operation Space Separation
We can segment or separate the space evenly or according the reach of the printheads, and each separated space can be printed without colliding. Yet, in reality the printheads mounts limit that operational space into slightly smaller spaces, but ideally:
nvolumes = volumetotal / volumeprinthead
If the individual printhead volumes aren’t regular, then we end up with arbitrary amount of printheads to cover a given print volume:
volumetotal = sum( volume1..n )
In reality, we require (slightly) overlapping operation to get seamless operation, so the “regular operation space separation” is only theoretically, but not practically.
Overlapping Operational Spaces
When the printheads can reach each other operational space, they become overlapping and controlling tool path generation needs to take care no collision is occurring (same place at the same time).
The Static Non-Overlapping Operation has static defined operational spaces where the operators can function – it’s quite obvious such solution is impractical, as in real life there would be space which cannot be reached, a kind of blind seam not reachable by either operator.
The Flexible Non-Overlapping Operation is flexible defined operational spaces, in the illustration above those spaces are co-dependent.
The Static Overlapping Operation is when those operational spaces are overlapping, yet, prefixed or static operational spaces.
The Flexible Overlapping Operation is flexible operational spaces, yet due the nature of the setup these operators cannot occupy the same space at the same time which would result in physical collisions.
Now, the last part of the last sentence may sound obvious to even mention, but bear in mind you can have two projectors shining light into a resin bath, and expose and solidify a 3D model, then these two lights acting as operators indeed occupy the same space at the same time as part of their function. So, the operators functioning with light can occupy the same space at the same time, whereas solid operators, such as robotic arms, cannot.
whereas parallelfactor is 1.0 if printheads can print parallel, or is less if the operational space is overlapping and preventing printheads to operate parallel thereby.
~ * ~
“Parallel 3D Printing / Additive Manufacturing” Part 2 follows later (will be linked when published)
Joshua Bird announced his 4-axis Polar CoreXB Printer December 1st 2024 on Reddit and YouTube:
Beside the hardware, he also implemented the required software and released it on github as “Radial Non-Planar Slicer” – an impressive undertake and success in one go.
4-Axis Linear Head(XZ), Rotational Head(Y)Bed(Z)
Even though I was impressed by the 4-axis RTN by ZHAW I wrote about, I abandoned the idea of 4-axis setup, as I considered it too limiting. Joshua’s approach to use a Z rotating bed, and the nozzle rotating ~180° or -90°.. +90° in Y-axis implements a polar (coordinate) printer.
He has a fixed Z-rotational center on the bed, and variable head Y-rotation in X. By keeping the Y-rotation (B) at 0° (vertical), the printer can print Z-planar cylindrical as well.
It can print XY Z-planar like a 3-axis 3D printer by mapping the XY coordinates into radial/cylindrical coordinates as well, so let’s summarize quickly what can be done:
polar coordinates, spherical slices
cylindrical coordinates, Z-planar cylindrical slices (fixed Y/B axis at 0°/vertical)
cartesian coordinates, mapping XY into radial/circular coordinates and fixed Y/B axis at 0°/vertical as well
So, this setup is quite powerful – and much better than the 4-axis RTN and also my 5-axis PAX setup, where the cabling and PTFE pipe for the filament prevent 360° Z-rotation / C-axis. Because Joshua’s has the Z/C-rotation on the bed, it’s fully continuous (or use a slip-ring to deliver power for a heated bed).
Dual Z Fixed X-Axis: Bad Idea?
Joshua’s original design is a cantilever, a single XZ beam with the Y/B rotation on the head. For a more large scale setup we require the X-rail to be attached on two Z rails, like the traditional Prusa-Mendel setup. In order to keep the -90°..+90° Y/B rotation of the head/nozzle, we need to position the rotation center a bit below the X-axis – but . . . this will prevent the nozzle to rotate Y/B to -90° or +90° as the gantry likely will crash with the already printed piece. So, it’s rather a bad idea to do so. As a result, we end up with a limited build volume, as the cantilever will tilt with further extending the X-axis.
Core XB: Combining Linear X with Rotational B (Y-Axis)
That is a very clever setup Joshua did, by having the linear X-axis combined with the rotational B-axis (Y-rotation) which he calls “Core XΘ” (Core X-Theta).
There are two motors (A & B) which control in “Core” manner the X-axis and the Y-rotation B-Axis or Θ
Non-Planar Polar Slicer
Joshua did also code a mixed non-planar slicer which combines multiple schemas as shown in the video:
Mixing spherical with free form slicing depending on the requirements of continuous printing without overhangs
That is the most tricky part, as polar (flexible tilt) and conical (fixed tilt) slicing may appear sophisticated, but it has its even more expanding challenges as Joshua mentions in his video, and obviously he developed a mixing strategy/schema slicing procedure to slice according the given requirements of printing continuous without supports – as when slicing non-planar, pieces otherwise easy to slice & print, suddenly have overhangs or are non-continuous or not connected anymore. So, he resolved it – let’s see the next weeks – will update this blog-post accordingly.
The S^3 DeformFDM Slicer from Manchester Uni (Tianyu Zhang, Tao Liu, Charlie C.L. Wang) has provided a solution to combine Support Free (SF), Strength Reinforcement (SR) and Surface Quality (SQ) under one umbrella of consideration and optimize all three requirements using quaterions per sample. I tried to adapt that solution but struggled to even build the solution from the sources back in 2022.
Impressive Step Forward
Joshua Bird has “single-handedly pushing the 3DP community forward.” according a YT-comment, and I whole-heartly agree, not only did he thought out of the box with the Core XB setup, but resolved multiple hardware and software challenges – impressive.
References
Polar Core XB 3D Printer (original YouTube video aka “I Built a 4 Axis 3D Printer Unlike Anything You’ve Seen”)
Formnext 2024 in Frankfurt/Main (Germany)Formnext 2024: Hammer ManFormnext 2024
Once again in fall (November 19-22, 2024) Formnext expo opened its doors, and I attended for 2 1/2 days – here my brief write-up and reflection of my experience.
Formlabs
Formlabs made an interesting move: it abandoned the SLA (laser-based) to MSLA (UV light & LCD masking) resin printing with their Form 4 series (Form 4 & Form 4L), so I briefly visited their booth:
FormlabsFormlabs: Form 4Formlabs: Form 4Formlabs: Form 4Formlabs: Form 4Formlabs: Form 4Formlabs: Form 4L
I highly recommend the taking apart of the Form 4 by Shane Wighton / Stuff Made Here to get to know all the engineering work which went into the new series.
The company behind the commercialization of the VORON series with their SOVOL printers, SOVOL 08 or SV08, a Core XY 350x350x450mm build volume, priced at 550 EUR.
So far only Genera‘s G3 does something similar where washing & curing is done in the same case/apparatus.
Prusa Research
Prusa Research announced at the Formnext their new Prusa Core One, a Core XY with 250x220x270mm build volume, nozzle temp. max 300°C, chamber temp. max 55°C, priced at EUR 1,350, with the ability to uprade from MK4S, available later in 2025.
Prusa ResearchPrusa ResearchPrusa Research: Prusa Pro Print FarmPrusa Research: Silver / White Iron ManPrusa Core OnePrusa Core OnePrusa Core One: DisplayPrusa Core One: Side with spool, nicely donePrusa Core One: BackPrusa Pro SLXPrusa Pro SLXPrusa Pro SLX
One may argue, Prusa reached a state of commercial success, that it can’t afford to support low-cost chinese replicas of their invention, exactly as it happened to Makerbot and Ultimaker – and both stopped innovate, and recently merged – let’s see if Prusa will thrive in hardware innovation the coming years or lay back as well.
Open Source Software seems more tolerable to commercial pressure: Prusa Slicer still actively developed and forked many times, and so the Cura Slicer by UltiMaker; both slicers enable countless university labs and commercial engineering offices to develop new hardware and having an Open Source implementation of reliable slicers with a ton of detail knowledge embedded when one studies the source code.
Rather unknown brand to me caught my attention the past year due to high praise from various YouTube reviewers: the older GKTwo (228x128x245mm @30um) ~670 EUR and the new GK3 Ultra (300x160x300mm @20x26um) ~1,200 EUR.
integrated heating system at 35°C
resin feeding system & resin weight measuring
built-in air filter
Industrial-level features while maintaining consumer pricing.
The booth featured Pengji (LCD maker) and Apex Maker (MSLA printers) together, so I assumed they are the same business with different brands with LCD making as core business – a representative at the booth confirmed my guess – so aside making UV LCDs they started to produce their own series of MSLA 3D printers:
Plus their own branded resins, curing and washing stations, see more at Apex-Maker.com
Elegoo
This year Elegoo released Mars 5 Ultra and the Saturn 4 Ultra which have tilting vat alike the Prusa SL1S which decreases layer print time (curing + detaching from film) down to apprx. 5 secs; so I was eyeing to get a Saturn 4 Ultra.
Thursday morning I revisited the booth and asked about whether Saturn 4 Ultra would permit to put a magnet flex plate on it – which would change the Z offset for homing, and two representatives explained to me it would be a bad idea as the home aka zero Z-position would fail, and one would require the change the “G-code” in the firmware (which is closed-source) . . . in other words, the Z offset or home position is somehow hard coded and doesn’t use pressure sensor to calibrate home (this Reddit thread shows Chitubox Slicer allows to alter the homing G-code, LycheeSlicer doesn’t have this option) – the usual rabbit hole.
My resin printing farm is composed by Anycubic printers (Mono 4K, X2 and 6Ks), mainly due the diverse third party market for replacement and add-ons, so I was checking out their new Photon Mono M7 Max (298x164x300mm @46um) priced at EUR 900 (2024/11).
I was hoping Anycubic would also implementing tilting vat to decrease the print time and allow more delicate prints due more gentle / gradual detaching from the film with the tilting motion – so far not yet.
This basic looking booth I consider a gem among a few others at Formnext 2024, a sample of massive piece which was extruded with aluminium below the melting point using Additive Friction Stir Deposition (AFSD) procedure and yet creating a fully solid piece where layers fully bond – their main photo (1st photo below) or illustration at the booth was deceptive small at first sight, but after looking twice, I realized the entire machine “Fooke AM50” is over ~6m long with a huge build volume of 6 x 3.5 x 1.5m, and aluminium, magnesium, copper, steel and titanium as possible materials, with build rates up to 15kg/h – and all extruded in normal room atmosphere and temperature.
Following video gives an impression of AFSD & FSW based extrusion by MELD (unfortunately the audio is quite silent in most parts):
grumpy, because of the patent collection which caused the 20 year delay for others to be able to adapt major extrusion technologies, once they expired the “3D printing era” actually started with affordable 3d printers (MakerBot, Ultimaker, Prusa, and all the chinese replicates) and many prosumer derivates
grandfather, because it’s truly the beginning of extrusion-based 3D printing / additive manufacturing
The cylindrical shape of a 3D printer caught my attention, and on the 2nd sight I saw it was a Positron-like setup of printing upside down, the Kokoni Sota printer:
I briefly spoke with Tianrun Chen and he explained when printing upside down, less support material required, and therefore developed their own support generation method which reduces material by 20-50%.
210x200x230mm build volume
up to 300°C nozzle temperature
optional multi material systems (also in cyndrical case)
National Taiwan University of Science & Technology (NTUST)
The little booth, and the walls where full of large scale posters with a lot of information, so I stopped and began to absorb the information – in particular the mention of TPMS caught my attention.
32″ LCD Resin Printer50″ LCD Resin PrinterResin 3D printed, heat activated foaming 2x in XYZ
So, one of their invention is a resin which when it is baked at 120°C it increases in size XYZ 2x – therefore in volume 8x – which they named “Foaming Resin Printing“, the sample prints like the bike helmet or even the small TPMS Schwarz P cube was incredible light compared to its size.
Another invention is printing TPMS like Schwarz P and then fill the partial closed P cells with a foaming agent, they call this process “Multi-Material Additive Manufacturing using Hybrid 3D Printing and Filling Process” (see non-free paper) – in case of the Schwarz P TPMS the cells are interconnected and spherical, so the agent flows into all captivities quickly as “Liquid Filled Closed Cell Lattice Structures” – and reduces print time while increase stiffness / strength – kind of combining Additive Manufacturing with Mold Injection method.
3D printing closed cells with injection of secondary material
Another achievement is their 32″ (697x392x500mm build volume) and 50″ (1,150x600x500mm build volume) large LCD/MSLA resin printers, with new film coating of better release of the cured layers, although we couldn’t get into the deep details of it. See this non-free paper and this paper for details about the 32″ resin printer.
LCD 32″ 3D PrinterLCD 50″ 3D Printer32″ LCD type vat photopolymerization 3D printing system
I could easily spent an hour to explore their different projects and the benefits of those 3 distinct inventions alone. See more at NTUST.edu.tw and its High Speed Printing research center, in particular follow Mayur Prajapati.
TUM (Technical University of Munich)
The TUM had a very prominent and large booth with many departments and many samples shown:
What caught my attention was the “Infinity Node” made from ~1mm gravel, it was printed using SLS-like printing: Particle-Bed 3D Printing by Selective Cement Paste Intrusion (SPI), using concrete as a binder to connect the small gravel together. The “Force-flow Optimized Node” (brown triangular truss shape) was done using Particle-Bed 3D printing by Selective Cement Activation (SCA).
The other samples were “Molten Metal Jetting” (Copper), “Functional Graded Material” using Plasma Directed Energy Deposition with Powder Feedstock, and “Laser Cladded TUM Logo” using Laser Metal Deposition (316L + Construction Steel).
Quantica caught my attention 2-3 years ago, also at the Formnext, as I saw the modular fully voxel-exact jetting of material. Their core business was their printhead “NovoJetTM” able to print high viscous materials, something other jetting printheads would not able to do.
This year I talked to Sven, and he explained to me that they broaden their application areas like exact glue jetting and other high viscous materials, as some industries saw their printheads as high precision material deposition, outside of the “3D printing / Additive Manufacturing” use cases – quite interesting how their core expertise found other applications.
Quantica Print Engine: circular motion of material kept under pressureNovoJet Printhead: 96 nozzles, 50um diameter each, 8kHz max frequency
Again I asked for some 3D printed samples, and again I was denied any like the past 2 years (!!) – they said they don’t have any (or very few) but just for internal use; my main interest is to observe the actual blend or non-blend and then 3D dithering of materials and colors on drop level in order to explore new applications.
An interesting approach, new to me, is Xolography, spatial printing in a resin – first a light sheet is established in the Z plane of 50um thickness, and then XY image is projected into the volume, different color/wavelength for each, and where the Z plane and the XY image light meet, there the photopolymerization takes place; it only works with highly transparent resins, more precisely dual-color photoinitiator (DCPIs), as otherwise the Z plane won’t reach all XY 2D image.
Xolography: UV light Z plane intersecting with XY plane with visible light
first wavelength (375 or 405 nm, UV light) activates the dormant initiator by triggering a spiropyran photoswitch (Z plane)
second visible wavelength (450-700 nm, visible light) excites the benzophenone component to initiate polymerization (XY plane)
So far they reached 50x89mm XY size, yet any size in Z including continuous Z printing.
They claim “no layer lines”, and told me they project a “video”, yet, a video has also n-amount of images e.g. a certain frame rate. What they rather mean is a continuous Z motion while they project their Z sliced XY images, and because the Z motion is continuous, the “layers” rather blend together and get smoothed out. They also don’t need any support structures, as the cured/solidified structure stays in-place and doesn’t sink.
Their requirement of highly transparent resin they derive one of their main use cases: 3D printed optics, and isotropic properties (angle independent strength, no deliminations), other resins they developed are bio-compatible and even soft and gel-like – quite an impressive achievement.
Currently they have two machines available:
Xube2: 10x18mm (@5um feature resolution) up to 50x89mm (@25um feature resolution) x80mm, apprx. 6mm/min in Z, feature size 5-25um as mentioned, although 50um lightsheet width/thickness, so the “feature size” more applies to XY than Z
Xell: 10x17x10mm build volume, @10um feature size
This process has been new to me as mentioned, but while researching I realized it has been around since 2020 or so, when this paper was published.
They showed a bio-attributed Cellulose Nano Fiber (CNF) made from cotton fibres, and then reinforced polyamide PA66 into a filament named CNF/ECONYL® Polymer with extended heat resistance, strength (performing better than carbon fiber), smoothness and formability – actual details of measurements are missing to backup those claims – according their announcement the material becomes available in Q3 2025.
While their booth has been very small and unassuming, the company has nearly 50,000 employees world-wide.
As far I saw they are one of few which makes resin printers which print, wash and cure in the same machine (Genera G3), where as their G1+F1 and G2+F2 they call “glove free process with shuttle technology and automated post-processing”, where the build plate is carried over in an encasing “glove free” to wash and cure station (F1 respectively F2).
Genera G1 + F1: 134 x 76 x 140mm build volume @70um pixel resolution
Genera G2 + F2: same as G1 plus automatic platform change for queued printing
153-384 x 87-216 x 320mm build volume (@40, 70 or 100um pixel resolution)
Genera G3: same as G2 but has F2 integrated, print, wash & cure in one machine
268-384 x 153-216 x 320mm build volume (@70, 100um pixel resolution)
Genera G3 (3D Model): (left-to-right) resin vat, washing vat, and curing vat
Printing metal poses the biggest challenge in my eyes as you have to put immense amount of energy to deform metal, while maintain or desire high degree of precision. So Meltio focuses on Directed Energy Deposition (DED) and Wire-Laser Metal Deposition (W-LMD): stainless steel, carbon steel, titanium alloys, nickel alloys, copper and aluminium:
Their sample prints are massive, and quite rough compared (0.8-1.2mm wire diameter) to Selective Laser Melting (SLM, powder-based) models where we get 50um feature size.
Their “Engine Blue Integration Kit” integrates into existing CNC machines making it hybrid operation and achieve traditional CNC tolerances and feature size – one of the many examples where Additive & Subtractive Manufacturing are put together with their respective strengths.
Fraunhofer is a prominent research institute in Germany, with multiple booths displaying cutting edge AM methods – I didn’t have the time to dig deeper into their samples, somehow nothing stood out prominently, or it was vaguely explained at first sight unlike at other booths.
Whenever I pass by their booth, the only thing I think is “car-sized 3D printers costing millions of EUR/USD”, it’s a different use-case from where I operate – as simple as that, and yet, still impressive what they achieved.
Lasertec 65 DED Hybrid: DED with CNC capability, with just 735x650x560mm build volume
Btw, its name comes from multiple mergers: DMG (Deckel, Maho, Gildemeister) and Japan’s Mori Seiki – a blend of german and japanese workmanship.
A few years back in 2017 they released their Mosaic Palette, splicing (cutting and melding together) filaments ahead for a single filament yet achieving multi-color with a single nozzle without material waste. Meanwhile they massively scaled up their expertise and developed their own printers called Mosaic Element and print farm called Mosaic Array – and I admired the overall case design in simplicity how the filament cartridges are attached.
Element HT2 specs:
nozzle up to 500°C
build plate up to 120°C
build chamber up to 80°C
350x350x350mm build volume
up to 8 materials via filament cartridges
price starting at EUR 9,500
Their print farm is built on stacked printers, and a robot arm removing the built-plate with the printed piece and stack it below and refill the printers with empty build plates again.
While roaming in the halls, I recognized a few YouTubers like Ross Graham (FauxHammer), Jon Schone (Proper Printing) and also Joel Telling (3D Printing Nerd):
Here his marathon live-stream at Formnext 2024:
Among others were also Xolo with the Xolography light printers, Spherene with Daniel Bachmann, Josef Prusa showing the Prusa Core One, and Stratasys representative featuring SAF H350 able to process used powder with High-Energy Absorption Fluid (HAF) with Selective Absorption Fusion (SAF).
Tony Lock (Duet) & Jon Schone (Proper Printing)
As I was visiting the booth of OST (Eastern Switzerland University of Applied Sciences) & Spherene I met Jon Schone (Proper Printing) and Tony Lock (Duet3D) and we had a brief talk on multi-axis firmware & slicing and parallel printing as Jon showed in his video with a multi-gantry setup – I’m currently composing/writing on a blog-post about parallel/concurrent printing with multiple nozzles so I was keen to briefly exchange thoughts on this topic.
Tony Lock (Duet3D) & Jon Schone (Proper Printing) – Formnext 2024
Reflection
Formnext is an overwhelming expo on all things on Additive Manufacturing, and this year less people and less exhibitors than last year. Also a few startups with big booths last year either had a small one (nTop, Inkbit, etc) or were no longer present. A couple of newcomers and startups, and some companies I have been following pushing their innovation further and exploring new applications, like Quantica.
I only stayed for 2 1/2 days (Tue-Thu) instead the full 4 days, as from last year I knew I wouldn’t be able to absorb more things with more time, there was only so much to take in. For me it was worth it, to reconnect with companies and individuals I know only online and via video calls.
The “3D Printing Hype” has peaked, the stock prices (state 2024/11/22) of Stratasys (8% or -92% of max), 3D Systems (8% or -92% of max), Desktop Metal (1.6% or -98.4% of max) or Nano Dimension (15% or -75% of max) have shown how much air and speculation was there, and now the sobering realism and use case are established. The same time MakerBot & Ultimaker merged, two exhausted 3D printer manufacturer without hardware innovation in the past years – yet Prusa, Creality, Elegoo, Anycubic still thrive on the pro- and consumer level with both FDM/FFF and MSLA resin printers, and BambuLab so successful that it caught Stratasys’s attention to sue them about patent infringements.
It’s clear to me, that Additive Manufacturing (AM) has its unique use-cases, whereas injection molds still dominate mass production due the high volume capacities and cost effective production. Additive Manufacturing may be a bland term (I like “3d printing” better), but there are still many possibilities to be explored, as Xolo or diverse university booths have shown – any kind of granular or liquid material is melted, extruded, jetted, cured or bound with an agent.
Hammer Man in Frankfurt/Main (Germany) – Symbol of a Maker
At Formnext 2023 I spent some unexpected time to discover a new class of procedural structure called “Spherene” (“sphere” + “graphene”), it’s a name as introduced by a company with the same name.
It’s main feature is isotropic (“all directions”) distribution of forces. Their service provides the creation of this structure based on:
density (ratio of material vs empty space), hence their term of Adaptive Density Minimal Surface (ADMS)
form
wall thickness
where all of them are freely definable in 3D space contained within an overall boundary. Their service “renders” a mesh which complies with such, like defining at some point a lower or higher density, and transits in 3D space from one to another.
Spherene Metamaterial in Simulation-Based DFAM: CDFAM NYC 2024 (Video Presentation)
Patent
My immediate impulse was to code the Spherene aside of the existing TPMS’s, but I realized their business core is the service of creating meshes based on their procedure as described in a patent:
at its core it describes 6 steps (abbreviations added for clarity)
creating envelope
creating density field
adaptive Voronoi tesselation (AVT),
1st skeleton graph (SG) associated to AVT (SG-AVT1)
generated from the edges of the Voronoi cells
2nd skeleton graph associated to SG-AVT1 (SG-AVT2)
generated using Delaunay tetrahedralization
minimal surface from SG-AVT1 and SG-AVT2, using equidistant from both skeleton graphs, with minimal wall thickness requirements
Abstract: A method of additively manufacturing a minimal surface structure of a three-dimensional article includes a computer executing the steps of recording, in the computer,
an envelope of the three-dimensional article; generating a density field across a volume enclosed by the envelope with densities of the density field corresponding to local requirement values of at least one physical parameter at respective positions of the three-dimensional article;
generating an adaptive Voronoi tessellation of the volume using the density field;
generating a first skeleton graph associated with the adaptive Voronoi tessellation;
generating a second skeleton graph associated with the first skeleton graph; and
generating a digital minimal surface model from the first and second skeleton graphs.
The method may further include a 3D printer additively manufacturing the minimal surface structure according to the digital minimal surface model.
I think if Spherene is truly as significant for Additive Manufacturing, and an essential invention, it has to move beyond the grip of a single company and its patents – time will tell.
Samples
Daniel Bachmann from Spherene Inc. kindly shared with me a few samples, 20x20x20mm cubes, and 20mm diameter spheres with Spherene infills, illustrating their properties:
Spherene 20mm, 15mm and 10mm sampleSpherene 40mm in white & grey resin (top row) and white, grey and black PLA/PLA+Spherene resin samplesSpherene resin samplesSpherene resin samples
A few support structures were required for the spherical samples, the cubic samples did not require such:
Lychee Slicer: spherical spherenes required support structure due overhangs
Additionally I printed a few cubic samples with FDM on my CoreXY Ashtar C without supports at 40x40x40mm scale.
Subtractive Manufacturing & Molding Usage
The structure cannot very well machined with subtractive manufacturing processes – or only if the piece is sub-divided so all indentations can be milled, and sequentially fused or welded again.
Another approach comes to my mind is to form dedicated bricks, e.g. for large scale application like a building, and have a limited kinds of bricks depending on their position and use case, and have molds to form those limited kinds in larger quantities.
In order to produce a mold one would inverse the original model, the negative volume, that would be produced using additive manufacturing and then produce lost-form casting molds, or highly simplify the form so one can remove the positive without destroying the mold.
Mathematician August Ferdinand Möbius described in his paper 1858 a shape, which later became known as “Moebius Strip”: a surface with only one side or surface and one edge:
In the physical one can take a strip and twist it once and tape both ends – when creating a 3D representation, one rotates a strip by 180° to a closing circle.
Examples
Original Moebius StripMoebius Strip with 2 stripsMoebius Strip with 3 stripsMoebius Strip with spokesMoebius Strip with 2 strips & spokesMoebius Strip with 3 stripts & spokes
3D Prints
I printed some of the models on MSLA resin printers with plenty of supports, one more flat and another series more vertical oriented which required more support structure and harder to remove.
Moebius Strip Meditation: multiple turns: 1, 2, 3, 4, 5, 12Moebius Strip Meditation: multiple stripsMoebius Strip Meditation: single turn, double turnMoebius Strip Meditation: single turn, single turn with multiple strips, both with spokes
2024/01/22: published without much reflection & conclusion as research is ongoing
2023/12/02: adding more examples and refining details
2023/10/22: start writeup
Introduction
While studying continuous fiber 3D printing and its main nature is to find ways to lay fiber without interruption. In order to refresh my memory I revisited the Lissajous forms, which until recently only knew in their 2D form, the swirling strings or lines – and now extending it into 3D as well.
The main idea is to realize how a line, string or fiber can be used to fill non-planar and circumvent a 3D structure and how angular shifting in Lissajous context affects such form.
3D Lissajous
angle: 0 .. 2pi or 0 .. 360°
p, n, m: 0 .. 1000, the amount of loops
phi0, phi1, phi2: the angular offsets 0 … 2pi or 0 .. 360°
X = sin(angle*p+phi0)*r
Y = sin(angle*n+phi1)*r
Z = sin(angle*m+phi2)*r
I did a lot of experimenting – I could post hundreds of forms – but let me focus on one a bit closer, which got my attention:
While playing with 3D Lissajous, I thought to adapt the cyclic nature, but apply it to a circle laying in the XY plane and then rotate in X axis, and Y axis as well, and optionally cyclic translation as well:
d: diameter
angle: 0 .. 2pi or 0 .. 360°
p: amount of loops as in X=sin(angle*p)*d/2, Y=cos(angle*p)*d/2
q: amount of X rotations: rotateX(angle*q)
r: amount of Y rotations: rotateY(angle*r)
Spherical Lissajous 12.23 with spreading struts
The model was printed with MSLA white resin at XYZ 50um resolution with 120mm diameter, with a few support structures near the bottom:
Spherical Lissajous 12.23 with spreading struts MSLA printed at 120mm diameterSpherical Lissajous 12.23 with spreading strutsSpherical Lissajous 12.23 with spreading struts closeup 1Spherical Lissajous 12.23 with spreading struts closeup 2Spherical Lissajous 12.23 with spreading struts closeup 3Various spherical Lissajous
Spherical Lissajous with Translations
Using the Spherical Lissajous and extend it slightly:
2024/11/30: Mono X 6Ks reversed engineered (UV LCD, GUI LCD)
2023/11/18: published
2023/11/11: printer ordered
2023/10/24: starting writeup
Introduction
Anycubic Photon Mono X 6Ks
After using Anycubic Photon Mono X2, and not that happy with it, I saw Anycubic released another resin printer:
build volume: 195.8 x 122.4 x 200 mm (WxDxH), same size as Mono X2:
reuse all its third party parts also compatible with X2: vat, FEP films etc
resolution: XY 34μm, Z 10-50μm
9.1″ display with 6K resolution (5760×3600) display
monochromatic LCD (hence “Mono”), faster printer due shorter exposure
affordable with EUR 250-280 (2023/11)
no network, only USB drive printing
The Photon Mono X2 with XY 48μm has not so well performed for me in regards of precise parts, as 100μm precision was not really delivered, either it was underexposing and fail on 1mm walls, or solid walls but thicker – somehow I could not find a good balance. The Photon Mono 4K performed more reliable for me and was able to print geometrically more reliable and correct with XY 35μm; so my hope is that the X 6Ks delivers with 34μm pixel size.
I ordered the Photon Mono X 6Ks in 2023/11 for EUR ~230 incl. shipment, and it arrived a few days later from a warehouse in Germany to Switzerland.
Anycubic’s Naming Convention & Communication
The X 6Ks is kind of successor of the X2 with the same body shape and build volume, but who cares of good naming? It could have been X2 6K simply. Aside, Anycubic seems the care little to provide customers the XY resolution anymore on info material or their web-site (2023/11) but only state XY build area and XY resolution, and announces 4K, 6K or whatever – you need to calculate the XY pixel size yourself . . .
Print Settings
retrieved from Anycubic 2023/11/11 as screenshot:
For Lychee Slicer (2023/11) I recommend following settings:
enable Two Stage Motion Control (TSMC)
Bottom Layer: 1-6 (1 for small pieces, 6 for larger pieces)
Transition Layers: 10 (for flat pieces reduce to 1)
Lift Distances: [1] 4mm, [2] 8mm
Lift Speeds: [1] 1mm/s, [2] 3mm/s or 60mm/min, 180mm/min
Here my settings:
Mono X 6Ks machine defaultsMono X 6Ks resin defaults (two stage lifting enabled)
Flex Build Plate
I’ve got a spring steel plate 202×128 mm from Aliexpress, and as I ordered two of them for X2 but I can reuse one for the X 6Ks as well:
magnetic base: 2.2mm thick, slightly extends, 203x129mm
steel plate: 0.5mm thick, exact 202x128mm, I roughed up the exposed surface with sandpaper to increase adhesion
so one has to compensate 2.5 – 3mm in Z height and adjust, speak extend, the Z-level probe. I just glued a black strip of paper, and then of course re-level the bed again afterwards.
Glued a piece of paper extending it ~2.5mm (I used black marker to darken the paper afterwards)
cost-effective for 196x122mm build area and 34μm pixel size
good third party market (X K6s parts are mostly compatible with X2, except UV LCD & mainboard)
Cons:
Mechanical Quality Control (QC):
build-plate has too much horizontal play (2+mm), the Mono X2 build-plate which has the same size fits tighter (see below for detailed photos)
grey/clear upper case is/was warped, doesn’t fit seamless off by 3-4mm bent inside on left & right; likely package long stored in wrong position or bad clear plastic deforming over time (remedy below “Cover Wedges”)
no WiFi
very basic firmware
poor quality SD stick, it usually won’t last a couple of weeks before it becomes unwritable or unreadable, but this time I couldn’t even backup the data, the stick failed right away; visit anycubic.com and download the slicer, handbook and test files to calibrate UV exposure times
Anycubic is known to invest very little into the firmware which is provided by Chitubox, given they are one of the biggest resin printer sellers, it affects 100,000s of customers.
Mono X2 build plate mount (likely extruded): solid, little playMono X 6Ks build platemount: simple metal sheet bent + plastic distance where screws enterMono X 6Ks build plate mount: horizontal playMono X 6Ks build plate mount: horizontal play
The build plate also has new a black plastic connector to the mount, an attempt to reduce cost perhaps. The Mono X2 build plate mount is much better in my opinion, more solid; the Mono X 6Ks build plate has more wiggle/play horizontally, therefore pieces printed you want definitely in the center of the UV LCD/vat, not on the side to avoid any wiggle or play while printing introducing imprecisions.
Cover Wedges
As the X K6s came with a warped cover, and with the X2 with the same case geometry it was wiggly everytime I put the cover over, I printed a few “wedges” which align the cover easily:
white wedges (I removed to replace with black wedges)10x black cover wedgescover slides over properly
Formlabs 3’s & EMake SLA printers use a laser beam which has 100μm in diameter, but it can be positioned 25μm exact, the latter was used to calculate the value
The UV pixels are non-square, the longer side was used to calculate the value
The XY area in SLA scales not as good as with MSLA, as the laser beam takes longer the more XY area (e.g. more pieces) need to be rendered per layer – so, MSLA is recommended for aiming fast parallel printing. Interestingly Formlabs 4 is now a MSLA as well (2024/04), they seem to have abandoned the SLA laser-based approach.
Anycubic Mono X 6Ks, X2 and 2x 4K
One of the main reason I choose Anycubic MSLA is the third party market for add-on’s, like flex build plate, vats, LCD replacement etc, and I also like the build plate mount with 4 screws for alignment; and I hoped to replace the firmware of the Mono 4K which did not happen as the Open Source variant is still incomplete (2023/11).
Another year, another November in Frankfurt (Germany) and Formnext – this is the main event of the year professionally for me. As I reside in Switzerland the travel is fairly easy and short and the 770 exhibitors in two halls (11 & 12) with two floors each is so overwhelming that even 4 days attending is not sufficient.
Day 1 (Tue, Nov 7): I spent an entire day to explore hall 12.1 alone, which turns out a good choice as it was a dense populated hall with many smaller companies
Day 2 (Wed, Nov 8): visiting with a client half of the day to review some of their possible competition, and then explore 12.0
Day 3 (Thu, Nov 9): some schedules meetings and then explore 11.0 and 11.1
Day 4 (Fri, Nov 10): revisiting 12.1 and 11.1 briefly, visiting with another client some selected booths to check products on display
I surely missed a few booths in 11.1 and 12.1 still; whereas 12.0 and 11.0 were more large scale industrial AM solutions, mixed with university and regional focused booths which I didn’t have time to explore in detail.
Personal Selection
I feature some companies according my personal professional interests:
Spherene (Math)
I made contact with Spherene before via LinkedIn but I realized I missed the point of what Spherene actually “invented”, at their booth Daniel Bachmann took the time to show me the features of their new class of minimal surface model and it was challenging for me to follow him despite of my own experience with Triply Periodic Minimal Surfaces (TPMS) – after apprx. 20 mins I realized the scope and some of their depth of their “invention”.
Spherene: sphere sample with variable porousitySpherene: resin printed rabbit with spherene infillSpherene: FFF/FDM printed bone replace with spherene infillSpherene: print samples (MSLA, FFF/FDM, SLM)
In essence, the sphere is used as a base form, and density, wall thickness and other features are processed in a localized manner, filling the space. The main result doing is optimizing a form to distribute inner/outer forces, e.g. the ends of the spheres are perpendicular to the surface providing ideal way to distribute them into a network of thin walled interconnected spheres providing isotropic (“all directions”) property.
The samples on display were printed with MSLA, SLA, FFF/FDM or SLM were indeed very strong in relation to the printed volume, e.g. the hallow rabbit printed with resin barely gave in when pushing on the thin outer perimeter – impressive.
Their approach is available as cloud-based GUI or as Grasshopper/Rhino plugin. The actual details of their procedure isn’t easily found but a patent (WO2020229692A1) by CEO Christian Waldvogel gives some idea.
Genera (DLP Resin)
There are many MSLA/SLA/DLP printer manufacturers, yet, I wasn’t aware of Genera and I was shown their system, an integrated workflow:
all resin vatshave a lid (only applies for G1/F1 combo but not their bigger machines), which are opened only within the machine
the finished prints (still on the plate) are moved in a box into the washing machine (without any person touching resin or the resin coated prints)
once automatically cleaned and post-cured, the prints are removed from the build plate manually
In essence one does not interact with resin directly, it’s all contained within the workflow – which I like a lot. They also provide wide selection of resins: hard, soft, rubbery, opaque, transparent/clear.
My idea has been to adapt some of their approach to make my own resin printing with Photon series (4K, X2 and X 6Ks); right now I also have multiple vats, and flex-plate, but moving the printed parts and washing them are still messy.
Quantica (Resin Jetting)
Last year I already visited the booth of Quantica, and so this year again. I asked earlier for printed samples, but they declined, and again this time . . . it is bizarre to see a machine actually able to print, and they don’t hand out samples, but I was told by January 2024 I might get some. This tells me a few things, the printed pieces are very sparse or not yet at the quality they want others to experience – some samples were on display, but sealed behind a glass box unable to have in my hand. So I guess now, they are expecting or already have better and more reliable printing results where the printed pieces match other similar printing processes.
QuanticaQuantica: NovoJet OpenQuantica: multi material samplesQuantica: multi material samplesQuantica: multi material samplesQuantica: multi material samples
I follow their development closely since ~2 years as I consider it very innovative to print with 7 different resin-based materials at the same time and able to fine-tune material properties on the voxel-level.
Duet3D (Open Source Hardware & Community Building)
UK-based Duet3D with its Duet boards and RepRapFirmware (RRF) is, as I wrote before, a beacon within the Open Source Hardware community – it isn’t just an example for other companies, and but also a great synergy provider, aiming to bring different individuals, groups and companies together.
Duet3D: with David’s backDuet3D: Voron 0 with Brandon Build’s variant of Open5X Duet3D 5-axis Open5X with tool changerThe many companies using Duet boardsDuet3D booth: Josef Prusa & Tony Lock – legends of 3D printing community
Brandon Builds’ Open 5X version was featured on a Voron 0, and a second machine also 5-axis setup with a tool changer.
Rapid Liquid Printing / RLP (Flexible Structures)
While roaming around a small booth of RLP caught also my attention, where a video was featured of a nozzle moving in a bed filled with silicon printing rubber, and other flexible material:
Rapid Liquid Print (RLP)RLP: multi color samplesRLP: white “finger” sampleRLP: area with flexible “fingers” and color gradientRLP: Flexible seat print with inflatable cushions
Reinforce 3D (Enhancement)
Another truly innovative approach combining and enhancing existing additive manufacturing processes was shown by Reinforce 3D:
using existing AM methods such as SLM, SLA, MSLA and even FFF/FDM to make models with thin walled tunnels and then
filling or rather pushing them with strains of carbon fibres along with resin into the tunnels
and thereby reinforcing free forms by keeping the result lightweight but incredible strong due the embedded carbon fibres
Reinforce 3D boothReinforce 3D: SLM and resin (MSLA) printed pieces carbon fiber reinforcedReinforce 3D: FDM printed and carbon fiber reinforcedReinforce 3D: FDM printed and carbon fiber reinforced
A very small but significant detail is, that you can print multiple parts on a smaller printer, but once you start to insert the bundles of carbon fiber those segments of pieces get combined in a strong assembly, as the aluminium skeleton shown above.
INo Trident – inductive hotend by Plasmics: fast heatup and cooldown / 3s from 20C to 220C, 10s from 220C to 150C
At the booth of Plasmics I looked at the inductive hotend and saw the heating up in a few seconds from 20C to 220C and cool-off in a small demo first hand.
The hot part of the nozzle looks like a needle, with little thermal mass, hence the fast heat and cooling-off time, and then surrounded by ceramics with the inductive coil on it.
The hotend incl. the controller board priced at EUR 400 is high for DIY enthusiasts but low for an industrial setup.
And UltiMaker (after Ultimaker & MakerBot merger) wasn’t present again; the consensus has been that BambuLab‘s printers have taken the higher quality consumer FFF/FDM printers market segment, and the air getting thinner for UltiMaker – at the same time they are doing a great service with the Open Source Cura slicer.
Random Impressions
Metafold: samplesStratasys: samplesPhotocentricPhotocentric: samplesPhotocentric: samplesPhotocentric: large scale statueHP: samplesPrusa ResearchPrusa Research Pro seriesSnapmakerSnapmaker with CNC drillSnapmakerSnapmaker: IDEXFormlabsAnisoprintAnisoprintAnisoprintAnisoprintE3DE3D: non-planar / belt printer nozzlesApexMakerApexMaker & PangJi resin printer LCDsSintratec: SLS printersSintratecSintratecNexa3D: SLA & SLS machinesNexa3DAiBuild: multi-axis slicingElegoo: printers, filaments, resinsFormFuturaBambu LabBambu LabDyze Design: water cooled large hotends
running Klipper on it, likely requires expansion board with additional USB ports, not yet tested; ArchLinux repo provides it though
any kind of quick experimental setup when Linux is a requirement, and Raspberry Pi perhaps overkill already, and ESP32 not powerful enough to run Linux*)
*) as of 2023/09 a few people working on Linux for ESP32-S3
A couple of years ago I used ESP8266 with Lua, and ESP32 with Lua and MicroPython, but the functionality was very limited, although it had WiFi built-in, but barely ran anything more complex or serious – the Milk-V Duo changes this, at the same pricing of USD 5.00-9.00; and I really looking forward to have just a small device running Linux, and competing with Raspberry Pi’s which either hardly were available or sold at 2-3x the announced price.
Milk-V Duo 64MB RAM (bare)Milk-V Duo 64MB RAM with USB-C connected and SD cardMilk-V Duo 64MB RAM (backside)Milk-V Duo 64MB RAM, CPI ID: “CVITEK CV1800B AA202500 T6N605 2303A”
and the newer 256MB RAM variant (released 2023/12):
Milk-V Duo 256MB RAM (frontside)Milk-V Duo 256MB RAM (backside)Milk-V Duo 256MB RAM, CPU ID: “SG2002 AA202300 TM5M48 2357A”
Specification & Features have been partially copied from Milkv.io (2023/09 & 2024/02):
– boots, but then fails with systemd (coredump) – no login possible
Linux BuildRoot Disk Image
The BuildRoot is a minimal custom Linux release which is meant for IoT developers who know what they want and need and select the features at building time and then get a disk image which contains then a set of features and applications. It’s not very user friendly, e.g. there is no package manager which allows to add new packages afterwards once the system is running.
Attach a SD card to your host, find out which device it became (Linux):
% lsblk
CAUTION: make sure /dev/sdX (replace X with proper letter) is the SD card and not your other disk(s) as copying the disk image will erase and replace the content of the device you choose with the following command:
then remove SD card from the host*), and insert it to the MilkV-Duo board, and power it on via USB-C.
*) in case you have the wrong /dev/sdX, or write to it if it’s not plugged in, you might struggle to write there again, simply do sudo rm /dev/sdX and then pull out and plugin the SD card again, and you should be able to write again to it.
Ubuntu 22.04 LTS Network: automatically starts it USB EthernetThe client sets the network: 192.168.42.0 and the “host” chooses 192.168.42.100 its own IP.
After ~15 seconds you should be able to login with ssh root@192.168.42.1 into your MilkV-Duo with passwd: milkv or you attach a UART to USB bridge like that:
Pin 16/TX: RX/white UART-USB
Pin 17/RX: TX/green UART-USB
Pin 18/GND: GND/black UART-USB
don’t connect 5V/red UART-USB
and use tio /dev/ttyUSB0 (or another number) under Linux to connect via serial port; also useful in case bootup is stuck after doing changes, and ssh isn’t possible anymore.
Note: you won’t need to solder the male connectors, I usually just insert them loosely and the cable bending gives sufficient connectivity for a brief login to fix things and then remove UART-USB cable again.
[root@milkv-duo]~# ls /bin
arch dmesg linux64 nuke sleep
ash dnsdomainname ln pidof stty
base32 dumpkmap login ping su
base64 echo ls pipe_progress sync
busybox egrep lsattr printenv tar
cat false mk_cmds ps touch
chattr fdflush mkdir pwd true
chgrp fgrep mknod resume umount
chmod getopt mktemp rm uname
chown grep more rmdir usleep
compile_et gunzip mount run-parts vi
cp gzip mountpoint sed watch
cpio hostname mt setarch zcat
date kill mv setpriv
dd link netstat setserial
df linux32 nice sh
/usr/bin
[root@milkv-duo]~# ls /usr/bin/
[ fold od tee
[[ free openvt telnet
ar fuser passwd test
ascii gcore paste tftp
awk gdb patch time
basename gdb-add-index pip top
bc head pip3 tr
bunzip2 hexdump pip3.9 traceroute
bzcat hexedit printf truncate
chrt hostid pyserial-miniterm ts
chvt htop pyserial-ports tty
cksum id python uniq
clear install python3 unix2dos
cmp ipcrm python3.9 unlink
crc32 ipcs readlink unlzma
crontab killall realpath unlzop
cut last renice unxz
cvi_pinmux less reset unzip
dbclient logger resize uptime
dc logname scp uudecode
deallocvt lsof seq uuencode
diff lspci setfattr vlock
dirname lsscsi setkeycodes w
dos2unix lsusb setsid wc
dropbearconvert lzcat sha1sum wget
dropbearkey lzma sha256sum which
du lzopcat sha3sum who
easy_install md5sum sha512sum whoami
easy_install-3.9 mesg shred xargs
eject microcom smtpd.py.9 xmlcatalog
env mkfifo sort xmllint
event_rpcgen.py mkpasswd ssh xmlwf
evtest nl strace xsltproc
expr nohup strace-log-merge xxd
factor nproc strings xz
fallocate nslookup svc xzcat
find ntpdate svok yes
flock ntptime tail
[root@milkv-duo]~# ls -1 /usr/bin/ | wc -l
151
Multiple Milk-V Duos / Alternative IPs
In order to support multiple Milk-V Duos on the same host via USB-C, you assign for each board its own network:
board 1: 192.168.51.0
board 2: 192.168.52.0
board 3: 192.168.53.0
Edit on each board two files:
/mnt/system/usb-rndis.sh (buildroot-based) or /etc/usb-rndis.sh (other systems):
ifconfig usb0 192.168.51.1
/etc/dnsmasq.conf:
dhcp-range=192.168.51.2,192.168.51.242,1h
In order to add a new board, you login into 192.168.42.1 as usual, and change it to the 192.168.52.1 and so on.
Resizing Disk
By default the entire available space of the SD card is only 1GB (or 2GB in case you use another distro), but you can make the rest of the SD card available to /data for example – part of the guide was taken from a post in the forum but updated it:
% mkdir /data
% fdisk /dev/mmcblk0
n (new partition)
p (primary partition)
4
<RETURN> (confirm start selection)
<RETURN> (confirm end selection)
w
q
% reboot
I followed this guide to get ArchLinux working – thanks to Judehahh doing the main work – and added RNDIS support (Virtual Ethernet over USB) and made a disk image to use, the date e.g. “2023-10-09” references the riscv64 rootfs date which was unpacked as a base, the “x.xgb” describes the size of disk or rootfs and “vX.X.X” the actual release version. Unzip downloaded image first before writing on the SD card.
– 240MB RAM (no camera support) – 250MB swap enabled – RNDIS (connect via usb virtual ether) – minimum 8GB SD card – 5.7GBfree in rootfs – persistent MAC addresses for RNDIS (internet routing ready)
For the 256m-version disk image I downloaded the kernel from this thread and just copied boot.sd and fip.bin into the first partition, otherwise no changes toward the milkv-duo disk-image (below) was made
– 55MB RAM (no camera support) – 250MB swap enabled – RNDIS (connect via usb virtual ether) – minimum 8GB SD card – 5.7GBfree in rootfs – persistent MAC addresses for RNDIS (internet routing ready)
– 55MB RAM (no camera support) – 250MB swap enabled – RNDIS (connect via usb virtual ether) – minimum 2GB SD card – 960 apps in /usr/bin/ – only 40MB free in rootfs (!!)
– package management (apk) – rndis / usb network (but random MAC addresses) – rootfs 1GB, 150MB used
– dropbear v2022.83
– use apk [add|update] –no-check-certificate
For the 256m-version disk image I downloaded the kernel from this thread and just copied boot.sd and fip.bin into the first partition, otherwise no changes toward the milkv-duo disk-image (below) was made.
– package management (apk) – rndis / usb network (but random MAC addresses) – rootfs 1GB, 150MB used
– dropbear v2022.83
– use apk [add|update] –no-check-certificate
Notes:
use apk [add|update] --no-check-certificate to install new packages, otherwise installs or update fail
Ubuntu Disk Image
I followed this guide (see discussion thread as well) – thanks to Bassusteur – and added RNDIS related services so you can login with ssh root@192.168.42.1 (passwd milkv) via virtual Ethernet over USB; unzip disk image before you write on the SD card.
Notes:
apt / apt-get are awfully slow on Milk-V Duo at step “Building dependency tree...“, takes 4+mins for each apt install call as apt requires 50+MB RAM to build that dependency tree, which goes hard on all available RAM + swap
these are very experimental disk images
Note: milkv-duo (64MB RAM) vs milkv-duo-256m (256MB RAM)
– 240MB RAM (no camera support) – RNDIS (connect via usb virtual ether) – persistent MAC addresses for RNDIS (internet routing ready) – apt works now smoothly – minimum 8GB SD card – 6.0GBfree in rootfs
– dropbear (v2020.81) – python 3.10.12
For the 256m-version disk image I downloaded the kernel from this thread and just copied boot.sd and fip.bin into the first partition, otherwise no changes toward the milkv-duo disk-image (below) was made.
– 55MB RAM (no camera support) – 250MB swap enabled – RNDIS (connect via usb virtual ether) – persistent MAC addresses for RNDIS (internet routing ready) – minimum 8GB SD card – 6.0GBfree in rootfs
– dropbear (v2020.81) – python 3.10.12 – zram enabled in kernel, and works now, but apt still very slow
– 55MB RAM (no camera support) – 250MB swap enabled – RNDIS (connect via usb virtual ether) – persistent MAC addresses for RNDIS (internet routing ready) – minimum 8GB SD card – 6.0GBfree in rootfs
– dropbear (v2020.81) – python 3.10.12 – zram not enabled in kernel yet
– 55MB RAM (no camera support) – 250MB swap enabled – RNDIS (connect via usb virtual ether) – persistent MAC addresses for RNDIS (internet routing ready) – minimum 8GB SD card – 6.0GBfree in rootfs
disabled systemd-resolved service due clash with dnsmasq
added /etc/resolv.conf.tail with default DNS servers
installed dropbear (lightweight sshd), removed the auto generated keys, added in /etc/default/dropbear the -R switch so new keys are generated at first boot
Tips & Examples
Most of the examples relate to the Duo BuildRoot SDK setup, but should be easily adaptable to ArchLinux or other distros.
blink.py with sysfs GPIO
There is a way to control GPIO via sysfs with Python:
import time
import gpio as GPIO
pin = 440
GPIO.setup(pin, GPIO.OUT)
while True:
GPIO.output(pin, GPIO.HIGH)
time.sleep(1.0)
GPIO.output(pin, GPIO.LOW)
time.sleep(1.0)
and run it:
% python blink.py
See this table for GPIO names, pins and numbers, a copy (2023/10/10):
GPIO NAME
GPIO PIN
GPIO Number
Notes
GPIOA14
19
494
GPIOA15
20
495
GPIOA16
16
496
GPIOA17
17
497
GPIOA22
24
502
GPIOA23
21
503
GPIOA24
22
504
GPIOA25
25
505
GPIOA26
27
506
GPIOA27
26
507
GPIOA28
1
508
GPIOA29
2
509
GPIOC9
14
425
1.8V
GPIOC10
15
426
1.8V
PWR_GPIO4
29
356
1.8V
PWR_GPIO18
12
370
PWR_GPIO19
6
371
PWR_GPIO20
7
372
PWR_GPIO21
11
373
PWR_GPIO22
10
374
PWR_GPIO23
9
375
PWR_GPIO25
5
377
PWR_GPIO26
4
378
Pinpong
Follow the guide to install pinpong.zip (my mirror), only for V1.0.4 system image, and then:
blink2.py
import time
from pinpong.board import Board,Pin
Board("MILKV-DUO").begin()
led = Pin(Pin.D0, Pin.OUT)
while True:
led.value(1)
time.sleep(1)
led.value(0)
time.sleep(1)
The pinpong library covers quite a lot of functionality, and useful examples:
According this post, tinycc has been ported as well – C compiler and C interpreter in one – download the .zip (my mirror) and run its install.sh, and then fix missing executable bit:
lua (5.4.6): segmentation fault, luac (5.4.6): seems to work (both compiled with tinycc)
lua & luac (5.3.6) works
lua & luac (5.4.6) works
extras
quickjs/qjs, micropython, nano, screen, git, make (no gcc/cc, use tinycc), thttpd, nginx, lighttpd, php-cgi, file, which, sudo
pacman (package mgr), lighttpd, file, which, sudo, make
Internet Access for Milk-V Duo
RNDIS (Virtual Ethernet over USB)
The host has to run Ethernet over USB and also operate as transparent router and let the connected board(s) reach the internet, see this guide (use google translate to english), here the brief description:
On The Host
The outgoing_if is the outgoing interface, either eth0 or wpl0s0 or something, check with ifconfig of the proper name, and then as root perform:
Also, find out which IP your host got (ip_of_host) from the connected board, e.g. 192.168.42.120, also check with ifconfig.
On The Board
% ip r add default viaip_of_host
% echo "nameserver 8.8.8.8" >> /etc/resolv.conf
Static IP for Host with RNDIS
What may look simple actually isn’t that easy as RNDIS itself is the culprit:
one can limit the IP range in dnsmasq.conf to from/to be the same IP, but
as RNDIS assigns random MAC addresses to the RNDIS host (the board) and the RNDIS client (the host the board is connected to) treat it as new device at every boot – if you force it to have the same IP again, the host will not get a new IP via DHCP as it has the IP remembered for another MAC address . . .
it’s a mess
So, a working solution is:
/etc/rndis-macs.sh which generates two random MAC addresses but keeps them persistent then:
#!/bin/bash
RNDIS_USB="/tmp/usb/usb_gadget/cvitek/functions/rndis.usb0"
MAC_FILE="/etc/rndis-macs.conf"
generate_random_mac() {
printf "02:%02x:%02x:%02x:%02x:%02x\n" $((RANDOM%256)) $((RANDOM%256)) $((RANDOM%256)) $((RANDOM%256)) $((RANDOM%256))
}
# check if the MAC address file exists and has exactly two lines
if [[ -f "$MAC_FILE" ]] && [[ $(wc -l < "$MAC_FILE") -eq 2 ]]; then
# read the two MAC addresses from the file
IFS=$'\n' read -d '' -r -a macs < "$MAC_FILE"
dev="${macs[0]}"
host="${macs[1]}"
echo "using existing MAC addresses:"
else
# generate two new MAC addresses and store them in the file
echo "generating new MAC addresses:"
dev=$(generate_random_mac)
host=$(generate_random_mac)
echo "$dev" > "$MAC_FILE"
echo "$host" >> "$MAC_FILE"
fi
echo "dev_addr: $dev"
echo "host_addr: $host"
echo "$dev" > "$RNDIS_USB"/dev_addr
echo "$host" > "$RNDIS_USB"/host_addr
/etc/usb-rndis.sh:
/etc/uhubon.sh device >> /tmp/rndis.log 2>&1
/etc/run_usb.sh probe rndis >> /tmp/rndis.log 2>&1
/etc/rndis-macs.sh >> /tmp/rndis.log 2>&1
/etc/run_usb.sh start rndis >> /tmp/rndis.log 2>&1
sleep 0.5
ip link set dev usb0 up
ip a add 192.168.42.1/24 dev usb0
ip r add default via 192.168.42.2
sleep 0.5
systemctl start dnsmasq
Note: if you had 192.168.42.2 once assigned to the host, choose another “static” IP in your range, as your host has remembered the IP to a particular MAC address and won’t accept it again, e.g. 192.168.42.200 or so.
Here my brief guide – see also this guide (use google translate) – how to customize packages included in the base distribution of the image for the SD card:
% cd duo-buildroot-sdk/buildroot-2021.05% make menuconfig
then go into the “Target packages”, and then walk through:
Audio and video applications
Compressors and decompressors
Debugging, profiling and benchmark
Development tools
Filesystem and flash utilities
Fonts, cursors, icons, sounds and themes
Games
Graphic libraries and applications (graphic/text)
Hardware handling
Interpreter languages and scripting
Libraries
Mail
Miscellaneous
Networking applications
Package managers
Real-Time
Security
Shell and utilities
System tools
Text editors and viewers
once you selected the packages you like to have included, choose “Save” and confirm as ‘.config’ and then “Exit”.
% cp .configconfigs/milkv_duo_musl_riscv64_defconfig
% cd ..
% ./build_milkv.sh
and after while, depending on how many packages you selected, you find in out/ folder your new disk image you can copy on the SD card.
Note: buildroot is quite a quirky package, e.g. when you select a package and make a build, later deselect a package, it will still be included – worse, if you commit a clean slate in output/, some packages might not fully build anymore – you have to go back to an earlier state of fewer packages, remake the build, and restart re-selecting new packages.
Postfixing missing .so file
As of 2023/10 v1.0.4 buildroot-2021.05 environment, there seems a problem regarding a missing shared library for some of the compiled apps (like qjs), you can fix this:
% cd /lib
% ln -s ld-musl-riscv64v0p7_xthead.so.1 ld-musl-riscv64.so.1
– lua, quickjs/qjs, micropython, nano, screen, git, make (no gcc/cc, use tinycc), thttpd2), nginx, lighttpd, php-cgi – 55MB RAM1) available – “fixed” image has .so lib-fix included
– lua, quickjs/qjs, micropython, nano, screen, git, make (no gcc/cc, use tinycc), lighttpd (doesn’t work yet out of the box) – 55MB RAM1) available
apprx. 55MB actual available RAM, no camera support, INO_SIZE=0
thttpd, nginx and lighttpd are all http-server, only use one, see /etc/init.d/ where those are started.
Ethernet Add-On
The easier way is to get proper networking is to add a real ethernet port and properly wire it to the ethernet router. There are several options & sources (2023/09):
I did a small case for the bare board without Ethernet or Extension board to be 3D printed:
Milk-V Duo (bare): bottom case & lidMilk-V Duo (bare): bottom caseMilk-V Duo case closedPrinted in white & black PLAwhite vs black case blue & red LEDs see-throughwhite PLA cases allow LEDs see throughwhite PLA (FFF/FDM), white & clear resin (MSLA)MilkV Duo’s running BuildRoot & ArchLinux
As the Milk-V Duo has the same width and depth (X/Y) as the Raspberry Pico but it’s a bit thicker so the some of the existing cases work only partially:
Raspberry Pi Pico Case by Botvinnik works perfectly, recommended, bottom/lid with 2 options: closed or with holes
Raspberry Pi Pico Case w Slider 3D Print Model by AndysTechGarage bottom part works, but lid doesn’t fit due the camera connector; you need to disconnect board from power in order to put in or out from the case
Raspberry Pi Pico Case by Kuma0055: has pins for holes which Milk-V Duo PCB doesn’t have, not recommended
As soon my Ethernet connectors and Extension boards arrived I will provide case variants (2023/10/05).
