>>5645 Thanks. Yes I think both hovershoes or some other form of motorized system could serve well as motive 'footwear' in the interim. And gyro-stabilization is also generally a good idea, if power-hungry. Ultimately, typical human bipedal locomotion is the goal, but in the meantime we'll settle for alternative approaches as needs be. >PIDs. So, for the uninitiate, PID is some sort of feedback control mechanism then?

Proportional Integral Derivative https://www.mathworks.com/videos/series/understanding-pid-control.html >part 1 https://www.youtube.com/watch?v=wkfEZmsQqiA

>>5667 Apologies, that page seems to be account-walled. Here's a playlist https://invidious.snopyta.org/playlist?list=PLn8PRpmsu08pQBgjxYFXSsODEF3Jqmm-y

>>5667 Turns out this guy Brian Douglas has his own independent YT channel. https://www.youtube.com/channel/UCq0imsn84ShAe9PBOFnoIrg

>>5671 Apologies, don't mean to bang on about it, but I'm finding more good resources as I dig along. https://engineeringmedia.com/resources

here we have an interesting mechanism https://www.youtube.com/watch?v=uv-Qp8p8Jqg

>>5930 That is interesting thanks. Is there a paper or anything more available on it Anon?

>>5931 nah this was random video in my youtube feed. I am wondering if this mechanism will be enough to achieve balance

>>5933 Here's his channel, seems like he's done a few of these. Amateur designer? https://www.youtube.com/channel/UC7hw6ztc_mnFh630R0MS5uQ

>>5936 https://www.youtube.com/watch?v=vlBgMg9WFfA men god. Luckily my family is heavy into machinery and maybe I will be able to build this. only the mechanism not the propulsion

>>5982 I think if you used 4 of those MIT Mini Cheetah clone thin disc actuators talked about here >>4890 in the hips, then you could probably pull this off. Apparently they run ~$350-400 ea. so plan on about $1500 for the actuators. But if you can get this working it would be both relatively inexpensive and remarkably efficient afaict. That anon's exact design would need some modification. The support frame for the hips would have to separate out the two pairs far enough apart to work well for the onahole cartridge inside robowaifus. https://biomimetics.mit.edu/

>related xpost >>8704

Feet development, related to DARA project >>8704

Are we not making this harder than it needs to be? For example there are reasons humans have a higher center of mass and it takes more power and effort (and years of practice) to walk: our need to be 5 or 6' tall fluid filled bags for one Lets take advantage of the clean slate we have with robo waifus. Several stacking advantages could make mobility a non-issue 1. Unfixed center of mass in design: as in "megaman" robots (and the Roll example), center of mass is drawn to the Boots, which themselves could be weighted further. (for "intimate" times the boots could be taken off to reveal more human-like feet, when sitting comfortably or lying down, etc). The rest of the structure doesn't need much, we don't need steel I beams for a skeleton, and the power battery and heavier parts could be set lower on the frame for the purpose of lowering the center of mass. 2. Gyroscopes: as in the Cube example above, gyros provide stability via angular momentum which could even be used to "push" against to maintain balance. Not sure how power consumptive gyros are but I imagine once they are set spinning it doesn't take a whole lot more to maintain the momentum with decent bearings and lubrication. 3. AI assisted learning. Same as how we learned to walk. Trial and error until a way to tackle moving from any point A to B across any slope or terrain is simply reflex. (I'd like to talk more about assigning movements to "reflex" CPUs, this is key since obviously we dont go thinking about every specific movement we need to make, we just think "go down stairs" or "open the door" and our habit and reflex handles the rest, correct?) Probably more to add to this list, but I see no reason these three advantages would not stack in a way to solve the problem of mobility simply and elegantly.

>>9053 The idea has been bounced around here of a Roll-like waifu, and certainly keeping track of center-of-gravity is vital for bipedal locomotion. As to your third point, particularly > (I'd like to talk more about assigning movements to "reflex" CPUs, this is key since obviously we dont go thinking about every specific movement we need to make, we just think "go down stairs" or "open the door" and our habit and reflex handles the rest, correct?) This seems to be pretty closely related to the idea of 'Neuromorphic Computing'. (You can find a paper in the Library thread). I'm inclined to agree with your ideas on this, and it's certainly good biomimetics, nervous system-wise. It entails it's own complexities too ofc, but I imagine the tradeoffs will be well worth it. Good ideas, thanks! It would be nice if you'd sketch some ideas out for us so we can see clearly what you mean about 'stacking advantages'?

>>9053 Some of this is food for thought. However, we're going to see different approaches, because people have different priorities. For example: I wouldn't want her legs to feel to light when I lift them, also her not needing to wear boots all the time. For outside, it would be okay though. Also, I don't get the problem until I tried it out myself. Humans do it, so it should be solvable. Another thing is, bipedal walking without help or guidance (walls etc) should be towards the end of any priority list, imho. For a domestic girlfriend it isn't very important. She could walk with help, using walls, being driven or drive herself with some device, walk on all fours, dancing while gripping something (e.g. a pole), standing up after grabbing something, then pose or do the dishes while keeping balance,...

>>9054 >Good ideas, thanks! It would be nice if you'd sketch some ideas out for us so we can see clearly what you mean about 'stacking advantages'? stacking as in multiplying, not merely Adding, each advantage Gyroscope + Low Center of Mass is an order more advantageous than Gyroscope alone or Low Center of Mass alone Because each is a different "vector" and not merely adding to the same vector Does that clarify things?

>>9055 noted, but also consider than a powerful enough internal gyro or set of gyros could effectively be an "internal" balance pole of sorts

>>9057 >Does that clarify things? Yes, it does thanks.

Some on topic videos about LOLA: https://youtu.be/Vzqm94bxjKI https://youtu.be/pmtKv8VEItY https://youtu.be/twfDcsQvNBY (Predictive Kinematics, Kinematic Optimization) https://youtu.be/piQm_oTYXIc https://github.com/am-lola

>>9672 Thanks kindly for the links, Anon.

>>9058 Agreed. IMO, any practical robowaifu walking system today -- given our current rudimentary state of kinematics sensing, planning, and animation -- will necessarily have some kind of gyroscopic stabilization to succeed. Here's an example of what good gyro systems can do >>5645 > the Cubli robot https://robohub.org/swiss-robots-cubli-a-cube-that-can-jump-up-balance-and-walk-across-your-desk/

Finally someone did a study on minimizing energy consumption in gaits. >We focus on the problem of developing efficient controllers for quadrupedal robots. Animals can actively switch gaits at different speeds to lower their energy consumption. >In this paper, we devise a hierarchical learning framework, in which distinctive locomotion gaits and natural gait transitions emerge automatically with a simple reward of energy minimization. >We use reinforcement learning to train a high-level gait policy that specifies the contact schedules of each foot, while the low-level Model Predictive Controller (MPC) optimizes the motor torques so that the robot can walk at a desired velocity using that gait pattern. >We test our learning framework on a quadruped robot and demonstrate automatic gait transitions, from walking to trotting and to fly-trotting, as the robot increases its speed up to 2.5m/s (5 body lengths/s). We show that the learned hierarchical controller consumes much less energy across a wide range of locomotion speed than baseline controllers. https://arxiv.org/pdf/2104.04644.pdf This could be adapted to bipedal walking by creating different expert policies for not only different speeds but also carrying items, moving hands and arms around, wearing a backpack, different size weights on the chest, cat girl tails and so on.

Adversarial Motion Priors for Stylized Physics-Based Character Control >Synthesizing graceful and life-like behaviors for physically simulated characters has been a fundamental challenge in computer animation. Data-driven methods that leverage motion tracking are a prominent class of techniques for producing high fidelity motions for a wide range of behaviors. However, the effectiveness of these tracking-based methods often hinges on carefully designed objective functions, and when applied to large and diverse motion datasets, these methods require significant additional machinery to select the appropriate motion for the character to track in a given scenario. >In this work, we propose to obviate the need to manually design imitation objectives and mechanisms for motion selection by utilizing a fully automated approach based on adversarial imitation learning. High-level task objectives that the character should perform can be specified by relatively simple reward functions, while the low-level style of the character's behaviors can be specified by a dataset of unstructured motion clips, without any explicit clip selection or sequencing. These motion clips are used to train an adversarial motion prior, which specifies style-rewards for training the character through reinforcement learning (RL). The adversarial RL procedure automatically selects which motion to perform, dynamically interpolating and generalizing from the dataset. >Our system produces high-quality motions that are comparable to those achieved by state-of-the-art tracking-based techniques, while also being able to easily accommodate large datasets of unstructured motion clips. Composition of disparate skills emerges automatically from the motion prior, without requiring a high-level motion planner or other task-specific annotations of the motion clips. We demonstrate the effectiveness of our framework on a diverse cast of complex simulated characters and a challenging suite of motor control tasks. https://xbpeng.github.io/projects/AMP/ Basically a way to use adversarial learning to imitate reference motions while seeking to solve a goal.

>>10160 >Finally someone did a study on minimizing energy consumption in gaits. Excellent. I consider that topic to be essential for finally achieving the idealized goal of a human-like robowaifu without having a state-sponsored R&D budget available. This will only just be on the very edge of what's even technically feasible at the moment.. If it's achieved at all, it will be due to everything on board the robowaifu being designed to be as efficient and economic as possible. Macro physical behaviors like walking will consume significant energy resources. So, not only will efficient gaiting be needed to even work in a natural lifelike way, but it should also make it possible to have our robowaifus walk around for more than 2-3 minutes without immediately needing to do a full 8-hour deep recharge before she can move for another 2-3 minutes, etc., etc. Ever wonder why a humanoid robot that's as well-funded as even Honda's ASIMO is never shown walking for more than just a few steps in any promotional? Because it's gait is just about as inefficient as it could be, power-consumption wise. And even significant steps forward in this arena such as Boston Dynamics parkour bot aren't shown going on long walks, in their case it's more a question of not keeping weights to a minimum, even though the gaiting algorithms are much better in that case. Every.thing. on board must be optimized for robowaifus to succeed and become the reality we all dream of. Everything.

>>10164 I think in real use robowaifus will be more like laptops, plugged in most of the time but able to disconnect and go at any time. My robowaifu would most likely be sitting 95% of the time and my place is so small she could get around while still being connected to power. It'll be an interesting problem navigating small spaces without knocking anything over and not burning excessive amounts of power to avoid stuff.

>>10181 >It'll be an interesting problem navigating small spaces without knocking anything over Yes, that definitely is going to require us to have an effective solution for at least two needs Anon: 1. Body-awareness. Distinct from Theory of Mind, this needs to encompass not only a detailed knowledge of her own physical volume & shape, but also a predictive methodology that can plan for her own kinematic/physics dynamics upcoming, based on her own short-term planned movements. 2. Excellent situational-awareness. She needs to have a detailed map of her environment at all times -- and one that can more or less instantly adapt to changes in it. Your own motions in the space for example, or your cat's. On a related point, until and unless we can actually give her a Theory of Mind, she'll have to rely on situational awareness (and possibly some types of crude heuristics) to be able figure out where you are/shortly will be. Thus (just one of the many) need(s) for a near-instantaneous observation capability.

>>10183 It would make sense to do this in a physics simulation. These were not only meant to train them, but to have one all the time for having awareness. Of course not simulating everything all the time, but when necessary. Also with simplified objects.

>>10186 Yes, that might be a good approach Anon, and for both of the enumerated needs too.

DrGuero made a new short video https://youtu.be/wxH3vQOz3JA, where he mentioned that his code is opensource now. His website http://ai2001.ifdef.jp is a bit slow right now, mb to much load after releasing the video. Also it's in Japanese, so some translation software will be necessary.

>>10343 Thanks Anon! Here's teh code page, I'm DLing now. http://ai2001.ifdef.jp/uvc/code.html >update OK, I've had a quick look at the code. It's old-style 'C with classes' (c.1985 C++) code. It seems to have only one major external dependency, Open Dynamics Engine. https://ode.org/ It's also W*ndows software. I'll try to sort out getting it to build on a modern C++ compiler on Linux sometime shortly, probably this coming week. Thanks Anon, he may be able to help our Bipedal Locomotion progress along.

>>10343 >>10344 update: OK, I've assembled the 4 code files into a workable project. I haven't set about trying to build it yet b/c dependencies. https://files.catbox.moe/wn95n1.7z However, if anyone would care to work on translating the embedded comments in the codefiles, then that would be a big help to me later on getting the code to run for us. TIA.

>>10344 >https://ode.org/ update By all appearances, the original ODE project has been abandoned. However, it's apparently continued under a new team. The project is now located at: https://bitbucket.org/odedevs/ode/src/master/ pamac info ode Name : ode Version : 0.16.1-1 Description : High performance library for simulating rigid body dynamics URL : https://bitbucket.org/odedevs/ode/ Licenses : BSD LGPL Repository : community Installed Size : 1.7 MB Packager : Antonio Rojas <arojas@archlinux.org> Build Date : 19/03/20 Install Date : 10/05/21 Install Reason : Explicitly installed Signatures : Yes I've gotten all the demos to build on Linux, so that's a good sign. I'll have to sort out all the dependency issues in the build file, but I'd think this project should be doable on Linux in the end. As is commonplace for me, I overestimated my abilities and it will be longer than I planned on. I'll post here when I have it all working. Probably be a week or two hopefully.

>>10409 Thanks, I would have helped already, but I have problems with my browser. Also there are a lot of comments, so I'll need to use sed or something like it. Great if you want to work on it, but bipedal walking isn't something very pressing. So don't worry about how long it takes.

>>10410 Thanks Anon, appreciated. I think r/n it's more of a thing to get working early, so that the AI geniuses here will have a readily-extensible tool they can use to work on connecting their AIs into, and coordinating a robowaifu's body movements, etc., via an early simulator environment. My personal goal for the software is to be able to continue refining it into an actual robowaifu runtime software, and for it to work onboard with small SBCs inside her. In my early tests thus far with ODE, it's looking pretty encouraging (~ sub-millisecond collision responses, etc.)

>>5409 Here the whole history of James Brutons bipedal walking robots, back to 2004 to around 2018. Robot X, shown in >>5409, was the end result. https://youtu.be/JWvH5PHKK74

>>10815 The real challenge isnt locomotion itself, its keeping the robot from falling over while moving. Though Boston Dynamics seems to have solved this issue. https://www.youtube.com/watch?v=fn3KWM1kuAw

>>11911 Everyone knows that they've solved it. Doesn't mean that we know how exactly or that we could replicate the same technique. You're missing the point. Bruton showed, what we can do as smaller developers or hobbyists. He didn't even study robotics. Also, he didn't overthink it, like some here tend to do. He just went on by try and error.

>>11911 >The real challenge isnt locomotion itself, its keeping the robot from falling over while moving. There's the rub, isn't it? >Though Boston Dynamics seems to have solved this issue. Now if only they'd just release their code open-source we'd all be good to go! :^) In the meantime, a Mobile Inverted Pendulum approach from UCSD has managed some verifiably-functional basic results (and we have direct access to the C code itself). Much more work needs to be done to mold it into human-like bipedal locomotion system suitable for a household robowaifu (>>7824). https://www.ucsdrobotics.org/mips https://github.com/beagleboard/librobotcontrol/blob/master/examples/src/rc_balance.c https://en.wikipedia.org/wiki/Inverted_pendulum

OK, so anon's OSRM project has gotten me back to thinking about light+strong robowaifu skeletons again. On the topic of bipedal locomotion a thought occurred to me I want to put out here: Since extra-dense, high-mass items like batteries are proportionally a higher percent of the total mass for a strawgirl or osrm robowaifu, wouldn't it stand to reason that mass could be articulated into a useful counterweight system within the lower torso/pelvis volume to swing back and forth as a type of counter-balance for the overall system during natural-gait walking/running/etc. ? The fact the mass is a higher percentage means it should be a particularly effective approach for either of these two types of robowaifus, but I imagine that it would be helpful even in a heavier robowaifu too. The basic idea is that you fix the batteries into an encasement attached to a short, inverted pendulum (inverted = batteries higher than pivot point). As the robowaifu's hips swing in one direction during her gait, the counterbalance pivots in the opposite direction. Animals already do this kind of thing instinctively during locomotion, and humans do as well. Our neuronal-musculo-skeletal complex all get wired up for complex kinematic motions such as these as we grow. By the time we're 5 or 6 yo, normally we pretty much have the 'force/counter-force' coordination down pat. Some of you may be aware that basically all skyscrapers constructed in any modern city have very massive 'floating' counterweights right at the top of the building that move this way and that depending on wind forces (and even earthquakes). A counterforce mass in our robowaifu's belly can serve a similar function and keep her center of gravity centered even when she's walking around. That is all.

>>12562 >Another Anon that actually understands physics and the importance of mass. Based and mass-shift pilled Check out Dr.Guero's work in this field. (Picrel) http://ai2001.ifdef.jp/mr1/mr1.html Femisapien also used this property. You can see her mass shifting point, it's the rotary encoder on her crotch. She shifts the mass of her body and electronics onto one leg, then moves the other leg forward as her weighted leg moves back. All you need is a parallel mechanism to keep her feet level and an added steering servo to twist her thighs for turning. Please do post any work you make even failures, I have a thread filled with my failures.

>>12582 Will do anon, thanks for link.

how soon we will have a bipedal robot that can imitate humans

>>13439 when the geniuses stop focusing on just the legs and finally figure out you cant have bipedal motion without fucking ears and start using gyroscopes

>>13447 How would you implement them?

>>13451 same way the human body does a feedback loop making continuous micro adjustments autopilots already do this with stabilizers, but thats easy for something with a plane perpendicular to gravity, parallel planes are in a league of their so dont bother until synthetic musclefibers become a thing

>Energy optimization during walking involves implicit processing [1] >Gait adaptations, in response to novel environments, devices or changes to the body, can be driven by the continuous optimization of energy expenditure. However, whether energy optimization involves implicit processing (occurring automatically and with minimal cognitive attention), explicit processing (occurring consciously with an attention-demanding strategy) or both in combination remains unclear. Here, we used a dual-task paradigm to probe the contributions of implicit and explicit processes in energy optimization during walking. To create our primary energy optimization task, we used lower-limb exoskeletons to shift people's energetically optimal step frequency to frequencies lower than normally preferred. Our secondary task, designed to draw explicit attention from the optimization task, was an auditory tone discrimination task. We found that adding this secondary task did not prevent energy optimization during walking; participants in our dual-task experiment adapted their step frequency toward the optima by an amount and at a rate similar to participants in our previous single-task experiment. We also found that performance on the tone discrimination task did not worsen when participants were adapting toward energy optima; accuracy scores and reaction times remained unchanged when the exoskeleton altered the energy optimal gaits. Survey responses suggest that dual-task participants were largely unaware of the changes they made to their gait during adaptation, whereas single-task participants were more aware of their gait changes yet did not leverage this explicit awareness to improve gait adaptation. Collectively, our results suggest that energy optimization involves implicit processing, allowing attentional resources to be directed toward other cognitive and motor objectives during walking. >Humans Are Designed to Think While Walking [2] >One of my favorite things to do when I am on vacation is hike in the mountains and take in as much scenery and contact with wildlife as possible. The former requires that I stay in good enough physical conditioning that I can achieve 15+ miles of mountain hiking per day. Therefore, when I am not on vacation, I go for a two-to-four-mile run every day before breakfast. That morning routine keeps me in physical shape and prepares me to undertake the research and writing projects for that day. Our problems are much, much simpler than God's were when He was designing us human beings with all our facilities, including these two. However, I'd say it's a good model for us to follow. After all, robowaifus should be able to talk with us about different things; like being young newlyweds while she's cooking a meal for us upstairs at the pub, right Anon? :^) Maybe Carver Mead's (et al) Neuromorphics can help us all out with this a bit. To wit: push the computation out to the edges of a [robowaifu's] system[s]. That way, while the 'autonomous' things are happening, her central-core computation mesh can be freed up to talk with us about important things. 1. https://pubmed.ncbi.nlm.nih.gov/34521117/ 2. https://reasons.org/explore/blogs/todays-new-reason-to-believe/humans-are-designed-to-think-while-walking >=== -minor fmt edit -add 'important things' cmnt

I think this general approach should be applicable, it is validated in real-life swiss quadruped robot [1] and in simulated [2] & real bipedals [3]. Compared to less-validated approaches it's a clear winner. You don't have to implement it in main AI, it's better if it runs on a small low-latency auxilary NN. It doesn't require too much compute or data, and the gait can be tuned via adding energy expenditure & smoothness terms to the loss. You can also include mocap data and tune the model on it for humanlike gait. 1. https://leggedrobotics.github.io/rl-blindloco/ https://www.youtube.com/watch?v=8sO7VS3q8d0 2. https://www.youtube.com/watch?v=hx_bgoTF7bs 3. https://techxplore.com/news/2021-04-robot.html