5-Axis Printing... Why Should You Want It? 

Thank you!

Thanks for the thought out reply! It is really amazing to me how complex of parts can be printed with how simple the layer by layer approach is. The layer by layer approach is elegant in terms of what can be accomplished with such a simple strategy. You are right of course, there is a market and demand for the technology, but the algorithms are very complex. I do think that there are likely strategies (intuition speaking here) that would allow for printing of generic shapes... the space of all potential moves needs some well reasoned logic to govern decisions about printed geometries and motion, preventing collisions, etc... and I obviously don't know what that would look like either. These algorithms wouldn't necessarily be optimal for all shapes or use-cases, but if there are a couple different general strategies to choose from, that would expand applications and usefulness drastically. I also think that even if the user gave some input to the slicing process, it would drastically reduce the needed complexity of the algorithms, too, which could be a possibility (hopefully something in-between current slicing and full blown 5-axis CAM).

I guess topological classification of objects could simplify the problem. Object's surface topology along with the size of printing toolhead are keys to deriving inner skeleton geometry. With a needle-like longer extruder, it might be possible to print inner skeleta (like trees) that have sharper edges (between branches), and then grow object from inside outwards. It'll come a time, when heuristics used by human-crafting 5-axis g-code will get reverse-engineered by training set for an AI to do it by providing pairs of target objects to print, with human crafted decompositions of them into tool paths, and with the rate of improvement in these fields make me think, such an AI will become available soon, and I hope it'll be open source, like alpha-fold is for proteins.

​@@IdeationGeek while heuristic-based AI implementaitons could have some promise, I'm not too sure myself. The immediately obvious problem is that the chic modern AI architectures that are getting the most attention with respect to really "hard" problems (things like LLMs, transformers, etc.) aren't exactly resource-efficient or all that smart. Most of the leaps and bounds we've seen with AI lately have been like a caveman carving intricate stone gears and levers to create machines; it's impressive, but the technology being used is still incredibly limiting. The recent llama models and mixtral are surprisingly competent for being something that can be practically self-hosted if you have a strong GPU, but they're still not great. I didn't recognize some Rust code and tried to get a quantized llama model to explain it to me, only to find out later that the snippet I was was looking at wasn't Rust at all, but Zig. Despite several back and fourths, the model never caught that, I naively said it was a snippet of Rust code and so that's what it kept saying it was. In other words, the popular models we're seeing now are built on some intrinsically flawed architectures that, while clearly capable of doing hard tasks, aren't likely to be favoured by consumers for practical applications. The other major problem is that I'm not entirely convinced efficient heuristics really even exist. Even reparameterizing incredibly simple geometries is something you only even begin to start doing in higher level calculus classes and it's quite involved. Outside of someone like Tesla with some weird case of synesthesia that has their brain wired to seemingly solve the problems magically, I'd be hesitant to believe anyone could make solid guesses for reparameterizations of arbitrary shapes and geometries. Maybe I'm wrong, but I'm not aware of anyone who I'd describe as "good" at doing something like that. (people obviously *_do_* do it, but I've only ever heard of it being a quite involved process, not really something that someone can get "good" at.) In short, I'm not entirely convinced a practically applicable general heuristic (or set of heuristics) actually exist(s). Again, it may just be an NP hard problem where there isn't anyway to meaningfully approximate better and better solutions in arbitrary cases and we'd be limited to using advanced lookup tables for known-good non-planar methodologies based on the localized surface. If this is the case, then heuristic optimization likely isn't ever going to be practically worthwhile. (at least not with classical computing, quantum systems though may still be able to pull it off) I'd love to see the technology come about, but it's one of those problems that I just can't hold my breath over. This feels like exactly the sort of problem that will *_always_* feel "just within reach", sorta like how nuclear fusion generators have been "another decade away" for like half a century. (though from what I've seen we actually have been making some progress on them lately so maybe not the best example) I'd be ecastic to be wrong on that and it is a solvable problem, but it's just hard for me to get invested personally. Every field has their white whales, material science has graphene, physics has a unified model of the universe, nuclear science has fusion power, etc. but when it comes to solving problems with computers it feels like so absurdly many problems end up just not being solvable efficiently. We've been able to brute-force through a lot of barriers thus far, but those have largely been due to technological/manufacturing advancements and we're begining to approach the theoretical limit for that sorta shit. (that's not to say there aren't improvements to be made, but it *_is_* to say that we need to accept that computers are a long-matured technology and the improvements are coming slower, faster. It used to be that computers lasted 2 or 3 years before the speed difference was so wild you just *_had_* to upgrade, now though you can go for easily double that before even thinking about an upgrade. The improvements are slowing down, and they're slowing down faster and faster as the technology gets more and more advanced) I won't say it's impossible, but I'm definitely far too cynical to hold out all that much hope for it personally. It very much feels like a "white whale technology" to me, for better or worse.

Very good point. Different printers with different abilities to rotate and clear a part will result in vastly different toolpaths with the same algorithm... to the point where that strategy could be ineffective or impossible, depending on the part.

Thanks Freddie! Thank you for your contributions!

Glad you liked it!

Thanks for commenting! Yes, it is amazing that the conventional layer by layer approach is so effective for such a simple algorithm. 5-axis and non-planar strategies add significant complexity, but also allow for many more possibilities with similar hardware (as like you said, you don't need extremely rigid systems to implement it).

Thank you so much! It means a lot to hear that. Excited to make more videos with even better production quality and content!

Many thanks!

Very much easier said than done. I am happy to see people starting to look into it though. Thanks for commenting!

Thanks! Even if you are printing a larger part, you don’t need a lot of contact area to “start” the print and then build on that section conformally to create the entire part. This is one way to mitigate a smaller build surface. Other than this, it could be redesigned to support larger build plates, as long as it is rigid enough for accurate prints.

A robot arm with an extruder would have more freedom of motion, but they are also very expensive! I don't know of a robot arm that would be applicable for 3D printing at this price point, and considering arms are serial manipulators with a lot of moving mass, I would think that it wouldn't be able to print as fast or with as good of a quality as a core XY motion platform.

Thanks for letting me know what you think! My apologies. I am working on getting a better mic and audio setup to make future videos sound better.

Are you suggesting to use a LLM to greate GC? I think this is an interesting idea, but considering that G-Code has certain constraints (like extrusion amount per mm travelled, a fixed geometry being a necessary output) makes me think that it wouldn't be a trivial application. Do you have ideas on how this might be implemented?

​@@Garth_42 I have experience 3d printing but not working with raw gcode yet so my thinking could be flawed. I was imagining something along the lines of defining desired print settings and some basic geometries at a system prompt level. Seems more realistic to try and have the systems apply existing "parts" in a well defined environment at this stage in AI progress. I also don't think it would be TOO difficult (for someone with more knowledge than me) to create some kind of low level validator, some system that attempts to confirm correct geometry/dimensions. A most simple example for this would be detecting the edges of a generated cube to confirm all lengths are equal.

My feelings exactly!

5-axis is available for subtractive applications, but is much less common for additive applications. Regardless, closed source or not, I am excited for advances in this field.

I don't see why rotary axes wouldn't be able to work with a delta style printer. An interesting aspect about deltas though is that getting to 6-axis is relatively simple... IIRC you "just" need to get the hardware to control all of the ball joints attached to the tool separately instead of 3 pairs of 2 ball joints being powered by the same linear guide.

@@Garth_42 could be interesting 🤔

I have no idea!

Thanks for the comment and the constructive criticism! Yes, I did my best to try to make it sound better in Davinci Resolve, but have literally zero experience. I have bought a new microphone since I recorded this video. I would appreciate a pointer to a video that you would recommend (if you have a recommendation and are willing to share).