Robot Head - Pro Version

Added: 2026-05-30

[00:00:02]I think that I can do better than this.

[00:00:03]In my last video, I went a bit overboard buying servos to try and make a servo tier list. And since now I have a big pile of decent hardware sitting around, I thought, why not build a pro version of my animatronic head? My open-source Alpha 2.0.33 head uses lowcost servos and common fasteners. But I wanted to see what would happen if I let go of all of these usual constraints and made something more powerful. One of the main things I learned from the last video is that server manufacturers don't adhere to any kind of standard dimensions. They're all over the place. So to start this project on the right foot, I decided to make a new branch in my on shape document and start using this software to its full potential by making the whole thing parametric. So if I choose a servo in the future, which is 1 mm wider, I just change one dimension, servo width, and the whole head reconfigures itself. And in fact, that's exactly what happened. I chose a faster, tougher, and smoother servo and finally broke out of 5V gel and started using an 8.4 volt servo.

[00:01:03]Now, that might not sound like a big difference, but I'm actually getting three times the torque and 1.5 times the speed. And that's really going to make the difference when it comes to stretching and posing the skin on the face. To make this work, I really had to think about the architecture of the design files which make up this head.

[00:01:20]This thing is really complicated. And I also wanted to design modules like the mouth, eyes, and brows to be modular.

[00:01:27]Firstly, I'm using On Shape's multi-art part studios. Here I'm using a single document to keep all of my variables and designs for essential repeating components. I can reference this document to create all of my modules, eyes, mouth, and brows. And then I bring these together to construct the face and finally the skull frame and later on the neck. Each step in this flowchart references the previous document. So if I make a change at a high level, all of these changes will be reflected in the documents downstream of the change. And that's how I'm able to manage so many complex interactions.

[00:02:02]As I was working on this, I was struck with the sudden realization. Our mouths don't move like this. The corners of our lips don't pull directly backwards like this. They move on a kind of diagonal plane. And I think that was a major issue with this old design. Now, deviating from the XYZ directions always makes your designs exponentially more complicated, but this was important. So, I moved all of the cheek servos to be in this new orientation.

[00:02:37]Before I got too deep into the mechanics, I needed to sort out the electronics on the old head. The wiring was a total mess. I tried to make each facial feature a separate discrete plug-and-play module which works mechanically but electrically it made things a lot more difficult. So this time I made a single heavyduty controller for all of the motors in the head. I really like the idea of having a dedicated chip to handle the servo orchestration, leaving a separate main processor free to do the heavy lifting, but making a single microcontroller output dozens of precise PWM signals simultaneously is just silly when dedicated hardware exists. So, I've come completely full circle all the way back to the PCA 9685, the exact same chip my very first IMAX used years ago. Except this time, I've designed my own custom board with two of them built right in and manufactured and assembled by JLCPCB. But what you ask is the spinning wheel for? I'll get back to you on that.

[00:03:35]Now that I'm in the process of getting my own custom servos spline adapters injection molded for my cogler project, it got me thinking, what other parts would I change if manufacturing constraints weren't an issue?

[00:03:49]To start with, I've never really liked heat set threaded inserts. I know that's a little bit controversial, but if I was building something out of wood, I'd use wood screws because the thread profile is designed to bite into wood. If I was using metal, I'd tap a thread or a machine screw. I always thought that heat set inserts are kind of like we're trying to pretend that it's metal, which is great if the part will eventually be metal, but this is plastic and it's going to stay plastic. I'm switching to proper plastic threading screws.

[00:04:18]The key is the 30° thread. This sharp acute angle displaces less material as it's driven, dramatically reducing the hoop stress, causes plastic parts to crack or split, especially when you're screwing in in a way which forces layer lines apart. I'm also saving a ton of weight because these are lighter than machine screws and I completely eliminate the need for the threaded insert.

[00:04:46]Last weekend, I tried to plaster a friend's ceiling. I've never plastered before and I thought it would be easy.

[00:04:52]Long story short, I wasted a Saturday and a couple of hundred pound. There's a lesson here. You can't be good at everything. So, for the first time, I decided to use somebody else's base mesh to design my silicone skin. Big thank you to Sako Sculpt because my sculpt always looked a bit goofy. On Shape has a really powerful feature where you can make a true mathematical CAD surface from a mesh file. And compared to how I tried to model it last time, this drastically cut down on the time it took and allowed me to make a much cleaner model and ultimately a better mold. This time I also overmolded the eyebrows and eyelashes in a darker silicone color.

[00:06:03]Apart from the few bubbles it left, I think this face looks a whole lot better than my previous attempt. In truth, I haven't spent much time with this yet, and I haven't experimented extensively with fixing it down, but I do feel a whole lot better about this whole system, much more so than the embedded magnets version. I saw a comment on an old video saying that I should use a full face keiny style skin that wraps entirely around the back of the head like a balaclava. And I thought they're right. It would work a lot better because it would maintain a nice even tension across all of the facial features. But guys, I really, really want to stay out of sex doll territory.

[00:06:40]I want this to look like a robot, not something else. So, I thought I could get the same effect while still keeping a robot aesthetic by clamping the outer edges of the skin evenly around the outside edges of the face, a bit like Cassandra. This is the two-part clamp that I designed with a flexible section so that it moves around the jaw. I also think I may have been overthinking a little bit how the skin attaches to the moving bits inside. Maybe glue isn't that scary after all. For the high stress points, I'm using a short shoulder screw inside of the silicone that just pops into the mechanism like a button cut. And I think that this paired with glue, paired with this kind of outer edge clamp is going to be enough to hold the skin and not infringe on any pattern.

[00:07:31]Now, partway through assembling, I realized that all of this mechanical design is effectively useless without an effective animation system. So far, I only programmed simple sequences. As I was assembling this new version, I made a feature for the new PCB which would enable me to calibrate each servo, saving its unique center points straight to a local JSON file on the hardware itself. And it remembers its own limits.

[00:07:54]And that's what the spinning wheel is for. It tells me which servo is being calibrated. And then I also have a separate pose mode where I can move the servo manually between its limit. It was very fun to code, but realistically this would be 10 times better as PC software.

[00:08:08]Never mind. Sometimes physical hardware is more fun. Defining the static resting points and the limits is easy. But if you want organic motion, it's not good enough. For example, to drive the motion of my eyelids in all of my standard eye mechanisms, the eyelids position is a function of the eyeballs position. And having these two different movement axes related to each other in this way makes the motion look 10 times more organic.

[00:08:31]For simple parts like the eyebrows, it's easy to say this is the bottom limit.

[00:08:35]This is the top limit. And no matter how I choose to drive that servo, whether it's a direct instruction or it's a mathematical function, I can tell it that it can never exceed either of these two limits. But some parts are a lot more complicated. So this is a single point driven by two servos. And we can use fairly simple trigonometry to work out the position of this point based on the angle of these two servos. And we can actually map that out on a 2D graph as a shape. And that shape will define all of the possible positions that that corner of the mouth can be without colliding into any of the other servos.

[00:09:09]The next layer of difficulty is that that shape will change when the jaw opens. When the mouth's open like this, we have much more space down here. And we may need the corner of the mouth to be lower for it to look more natural. So now when the jaw is open, we have a separate 2D shape which is completely different from the 2D shape we have with the jaw closed. Now we can map these two together and we can interpolate between them making a 3D graph. And that sounds complicated but it is doable. The real issue is that there are even more obstacles to avoid because we have these lip servos and depending on the position of these that is another factor which affects whereabouts this point can go.

[00:09:49]Now a 3D graph we could just about manage I think but a four or fivedimensional graph I don't think this is the right approach. My long-term goal for this project is to train an AI model to take the sound of speech audio and automatically translate it into physical facial expression. My idea was was to train the model on footage of speech using some kind of meshbased face tracking and then choose a vertex on that mesh which corresponds most closely to a particular servo in the design and then directly translate that to the servo here close to that point on the lip. And then it's just a onetoone mapping function. But while I was researching, I found out about an alternative way to do things, blend shapes. Essentially, these would be like preset facial poses. Instead of forcing the AI to micromanage every single servo independently, I just hardcode a few specific facial expressions into the robot, like an angry expression or a sad expression or mouth poses for different letters. And then the job of the AI model becomes way simpler. It would just listen to the speech audio and output a mixture of those presets. So rather than saying move server one 50°, servo 2 70°, it just says for this particular moment in time, look 30% angry and 10% surprised. So the servos will only ever be moving linearly between positions I have already hardcoded and physically tested. There's no risk of two motors trying to pull the same linkage in opposite directions and snapping the mechanism. But I admit I haven't fully wrapped my head around all of that specifics of how it's going to work. So there's still a lot of work to do, but now I feel like I have a really super solid mechanical platform and a robust CAD architecture in On Shape to start developing my machine learning model.

[00:11:39]This version is about 2/3 of the weight of this version. The servos are about 1.5 times faster and three times stronger, and I really think that will make the difference when it comes to organically posing the skin. If you're an engineer, I would really recommend checking out On Shape. Aside from all the stuff that makes them unique, version control, cloud-based interface, and collaboration, I do genuinely think it's just the best CAD design software at its core. You can get up to 6 months of their pro version entirely for free by using my link down in the description. My open-source robot head design is available for free in the description, but this pro version is a Patreon exclusive design for now.

[00:12:20]Joining my Patreon will give you access to my on shape library of all the designs I've created, including this brand new one and the servo driverboard.

[00:12:28]Thanks a lot for watching, guys. I'll see you in the next video.