Refactoring joints

[JoeJ]

Your joint limits are a bit off, e.g. you can not lift your knee so far up (you can, but you'd rotate your pelvis too to overcome the limit). But that's not important for now, you can work on balancing / working anyways. At some time i can help with screenshots from my own limits or i can tweak your numbers. (My own is not perfect yet either but much better.)

I wonder the twist / swing decomposition works for you with motor. For me it did not - too unstable. I assumed because rotation axis are not always orthogonal. If you run into problems that might be the reason.

I claim that to make the biped to keep it balance we do no nee to do any ray cast and there is not need to get a support plane.
I will let you think about that for a moment.

You mean user contacts and support polygon?
Yes i did balancing on one foot without all this, but if you want it as fast as possible you need the support polygon to calculate max accel, and when moving fast robust contacts seem necessary. (I did not implement user contacts for my ragdoll yet, just on the IP model so i'm not sure, and maybe your inverse stuff is more sensitive than my forward so avoiding the problem at all, but keep it in mind in case of issues. Without user contact my IP swings between 5 and 20 times correctly and then tips over for no obvious reason. The problem does not show up if it just keeps straight in balance.)

[Julio Jerez]

you are probably right, I add the anim tree pose, and try to move the leg to the standing up pose
but nothing is right, it jitter, the leg do no move to the right position and it is move a lithe is exposes.
no sure if it is because the joint.

[JoeJ]

Damn, but twist / swing decomposition is nothing else than using euler angles, so i guess that's an issue.

I suggest you look at my powered ragdoll joint and think about its requirement to know the rotation axis it will use - since that axis keeps pretty constant during motion it might work. (expressing limits would become more complicated but should be possible.)
May be worth a try before doubling the body count.

[Julio Jerez]

yes it comes down to be the same as an Euler decomposition. but I believe this is better that the quaternion decomposition.
I will post two image of a dummy moving his leg to the front and one doin a kick, you will see that the way the wing is don is using an Euler.
you are right the motion of the leg along the swing axis generates a twist torque, but I believe that is correct. but again I am not completely sure.

I believe there is some other bug some where, but can find it, I will probably have to simplify the test to a simple demo with 2 ands 3 dof and see why is wrong.

edit:
the fact that is exploding so each is a sign that the must be a sign wrong in the implementation of the joint, so I will check that out

I committed a repro case where it drops the biped fore 3 meters and is land find stable. bu if I do the inv dyin update is blows up so some how the inv is calaution wrong acceleration. so that a good sign.
anyway I will continue tonight.

[Julio Jerez]

I believe you are right, the twist acceleration is insane.
that has to be the bug.

[Julio Jerez]

ok I believe I found my bug, the join is totally wrong.
I wrote the order of the angular decomposition in teh wrong order, that why you see the parent body been slapped around instead of the child doing the twisting.
I will go over again.
I am tented to fuse the three joints in one so what I will do is that I will make a new one and if is work I will replace the other one at a time.

[JoeJ]

I am tented to fuse the three joints in one so what I will do is that I will make a new one and if is work I will replace the other one at a time.

Yeah i think it's not worth to have 2 and 3 dof, i use only one (and hinges), and i think i have no cases where one of the limits is unused.

[pHySiQuE]

Have you looked into using something like tensorflow so you can use AI for this?

[JoeJ]

Great video, very inspiring - thanks
Although it looks more like an animation system (did not read the paper yet) but totally relevant. I really like that idea of the raycasting worm - very nice kind of data for motion planning.

[Julio Jerez]

Yes, very cool indeed, but this is not what you think.
What that system does is that figure out blend factors, to blend a series of motion capture so that the resolution pose minimizes the error to the environment.
That does not generate motion, you have to have a system that play the animation already so that the neural network use as the goal.
Once the neural net is trained, and calculate the coefficient, the given the environment imputs, it will figure out the blend that when applied to the animations will generated the pose with the samples error.

If I am right the system we are working on should do it without the training, but I have been wrong so many time, that I have some douts.

Joe I thing I know what the bug is, iwillexplain later.
I was sorry this would cause some insurmountable problem, but now I know the was a reason for that behavior.

[pHySiQuE]

Here's one that is pure AI locomotion:

[Julio Jerez]

Those are the papers that I say fall in the category "Read the legal notes", and read them carefully.
I have not read that paper, but I am going to go on a ledge and predict, that after they self congratulate them self. In the conclusion or results it going to say some thing like. "We implement this using puts us library _______ and we runner at 2000 gets, maybe even higher, ......"

As if that part was just a foot note people should just ignore.
There is a reason why none of those Padres hasn't been adopted be the industry, they only have academic value, not real world applications.

The colleges giving this people masters, or even doctorates, at some point have to give then the degree after the students paid hundreds of thousands of dollars for the 30, 60, or 90 credit program.

This is in general the case, with few exceptions that come about every five to ten years when some one comes up with a truly new idea, and in today's world where information is the most valuable commodity, those people are snatch by scouts for companies like Google, apple, Qualcomm,Microsoft, etc.

I am so sure you the system I am envision will work, but I want to be useful, I am doing the best I can.

[Julio Jerez]

ok I downloaded the paper and this is what is says.

6 Experiments
WetestourframeworkusingOpenDynamicsEngine(ODE)[Smith 2006], version 0.8.2, using an integration time step of 0.0003s. We use the same ground contact model as Wang et al. [2009], simulating a spring-damper system with kp = 75000 and kd = 2000 throughuseofODE’sCFMandERPparameters(see[Smith2006] for details), and using a friction coefficient of µ = 10.

Time step is 0.0003 which meant the frequency is 3333, coefficient of friction 10, ...,
you can't say "we use the same friction model from some important dude" and in the same paragraph say that "we use a friction coefficient of 10", that's an argument from authority fallacy.
You tell me, is that acceptable physics?

That's my problem with people in pseudo academia, they are quick to dismiss right out of the bat newton saying that some how is slow to do box stacking of 1000 or 5000 boxes, but when they try other engines for what it really count they use these black magic number for tuning, and somehow it never comes up on their criticism.

I can guarantee you that if the three eggheads self appointed experts who wrote that paper had used newton, they'd get much better results at much lower frequency, maybe one tenth of what they used (300 hz) but they would never know because they read the opinions of people like: Kenneth Bodin, Adrian Boing and the other self appointed experts over that other physics engine forum and accepted as facts.

They had made a hit job straw man of newton and they keep doing it because they need to prevent people from even downloading the SDK and try the demos.

I know I mention this a lot, but the dishonesty of these people made my blood when I see this stuff.
They rather try to solve the physics shortcoming of the physics library using AI, as if artificial intelligence if going to replace the laws of physics) than admitting the physics library is physics in name only.

[Julio Jerez]

I am taking a step back and I decided to normalize the joints.

until now all joint are implementation their own policy to submit the row, as a result with the proliferation of joints, there are many inconsistences in the joint are submitted.

to formalized this I am making a robust 6DOF joint that is going to be use a s the base for almost all of the regular joints of the dCustom joint class.
this call will do all the house keeping of subliming the fix constraints and them call the sub class for issue the free dof like motors.

the join will also calculate the decomposition of the relative matrix into the three Euler angle so the all joint use the same consistent order unless the subclass decide to do a different order.

I believe I have working now and to test it I did the slider, and free 3dof, and the corkscrew, late I will do the hinge ( which is the work horse for many other joint things)

the result is that the callback is now very simple since it does not have to be concerned with row the define the joint behavior, it is only concerned with the free degree of freedom.

This also provide more robustness since the base class is use a lot more, therefore joint that are test much have less chance to generate malfunction bug when they are used. right now I see this happing with the some joint and my solution is to just make a new joint on the demo.

I believe this is worth the time before continue with the dynamic ragdoll.

[JoeJ]

I used a similar joint initially as well, with the rows aligned to the parent frame.

The issue here was that the motor speed was as expected if acceleration rotation axis was close to any parent axis, but much slower if the axis was off (e.g. 45 degrees). That's the reason why i finally aligned constraint space to motor acceleration.

That dates back 7 years now, but i suggest you test the motor behavior if it is too axis dependent.