/sci/ - Science & Math » Thread #11778978

84KiB, 1100x825, 5bc05727765453642d150312.jpg

View Same Google iqdb SauceNAO

Anonymous Tue 09 Jun 18:49:32 2020 No.11778978 View Reply Original Report

Quoted By: >>11779015 >>11779234 >>11779342

https://youtu.be/cTqVhcrilrE

>use neural net from video related
>run it "in the opposite direction" capturing motion data from humans and translating it to machine motion
>create robots that learn to behave naturally in the real world just by feeding them mocap data from humans

Can this be done? Is anyone working on this at present?

Anonymous

Anonymous Tue 09 Jun 2020 18:59:43 No.11779015 Report

Quoted By: >>11779026 >>11779116 >>11779281 >>11779342

>>11778978
>>Can this be done?
nope. It can only be used to make things that 'look' realistic. In animation you don't have to take forces into account, but in robots you do. It's also extremely difficult to abstract out forces from motion capture data. In addition, robots may not necessarily have the same dynamics as humans, so what works for humans might not work for robots.

Anonymous

Anonymous Tue 09 Jun 2020 19:03:37 No.11779026 Report

Quoted By: >>11779045

>>11779015
This should be able to be done by adding the transformation algorithm form earthly forces. OP is right. It could make entire characters run on their own in a virtual world. "Give them souls."

Anonymous

Anonymous Tue 09 Jun 2020 19:08:15 No.11779045 Report

Quoted By: >>11779054 >>11779116

>>11779026
uhh... ok. As I said it's very difficult to get force data from motion capture data.

Anonymous

View Same Google iqdb SauceNAO G013M.gif, 867KiB, 280x158

Anonymous Tue 09 Jun 2020 19:11:52 No.11779054 Report

Quoted By: >>11779100 >>11779116

>>11779045
Yes, but force data from instruments is plenty and open source. All it needs is that. Then bam. Instant modelling that can be overclocked and implemented in wider environments. The speed at which it can learn will be on par if not faster than an average rock climber.

https://www.youtube.com/watch?v=zVYknCLSx34

Anonymous

View Same Google iqdb SauceNAO projectgolem.jpg, 215KiB, 1275x750

Anonymous Tue 09 Jun 2020 19:25:24 No.11779100 Report

Quoted By: >>11779116 >>11779156 >>11779342

>>11779054
>>force data from instruments is plenty and open source
where? And force data applicable to humans might not apply to robots which will have different dynamics. Force data collected in ideal environments might not apply to real world environments.
>>Instant modelling that can be overclocked and implemented in wider environments
all I see is buzzwords, buzzwords, buzzwords...
I think you mean to say you're going to do learning in simulation, but the truth is our simulations suck. It's difficult to get things like friction right in simulation. So pic related is a classic example of this. Pic related is a robot that was evolved to walk in a simulation, but it only works on sand in the real world because sand has the closest properties of what the simulation used.

Anonymous

Anonymous Tue 09 Jun 2020 19:30:48 No.11779116 Report

Quoted By: >>11779156

>>11779015
>>11779045
great answer thanks
>>11779054
This is what i was thinking too. I don't see why you couldn't make force data from machines + computer vision data from mocap work together? not in the field myself so what do i know.

>>11779100
>where? And force data applicable to humans might not apply to robots which will have different dynamics. Force data collected in ideal environments might not apply to real world environments.
>where? And force data applicable to humans might not apply to robots which will have different dynamics. Force data collected in ideal environments might not apply to real world environments.
Interesting. But not impossible no?

Anonymous

Anonymous Tue 09 Jun 2020 19:41:41 No.11779156 Report

Quoted By: >>11779182 >>11779211

>>11779116
>>11779100
>>11779116

He's right about the friction variants, but you can offset that by gaining approximations of the terrain through visual sensors.
>running on ice, running on sand, running on rubber, running on grass, tall grass etc etc
Plenty of ways around that. Then it's just gravity, center of gravity, individual actions, borrowing leverage etc etc.

Anonymous

Anonymous Tue 09 Jun 2020 19:48:56 No.11779182 Report

Quoted By: >>11779211

>>11779156
It's faster to just upload all friction coefficients and have it run on them until it succeeds.

Anonymous

Anonymous Tue 09 Jun 2020 19:54:34 No.11779211 Report

Quoted By: >>11779230

>>11779156
>>approximations
and if you're off by 10% you tend to fall over.
>> just gravity, center of gravity
and if you don't know someone's weight then it's difficult to figure out forces. That being said, out of all the problems in robotics we're doing pretty well at making robots walk. One of the most difficult problems in robotics right now is getting robots to manipulate things. Motion capture won't help with that.
>>11779182
>>upload all friction coefficients
lol. It's not just friction, it's also how springy the terrain is, damping caused by the terrain, and more. But even if you have a policy that works on all friction coefficients, you still run into stuff like knowing not to step on a loose T-shirt on a hardwood floor because that could slip you or avoiding getting tangled up in some wires.

Anonymous

Anonymous Tue 09 Jun 2020 19:58:54 No.11779230 Report

Quoted By:

>>11779211
>and if you're off by 10% you tend to fall over
To be fair, that happens even to humans that aren't aware of what they are standing on
>springy terrain
Just another paradigm for the function.

Anonymous

View Same Google iqdb SauceNAO 65750956-783D-4142-BED8-F24A7ACD (...).png, 38KiB, 600x800

Anonymous Tue 09 Jun 2020 19:59:57 No.11779234 Report

Quoted By:

>>11778978
>neural net

Anonymous

Anonymous Tue 09 Jun 2020 20:11:56 No.11779281 Report

Quoted By: >>11779315

>>11779015
>equirement that reviewers are expert in the topic of the essay rather than experts in the field or focus of the journal
>>Established an interim editorial decision-making structure, effective immediately (read more about this newly appointed team here)
>>Removed the article from the journal’s website and launched an internal review according to the guidelines of the Committee on Publication Ethics (COPE)
>>We recognize and take very seriously the role we play in the academic system. Today, we also pledge to:
Some animation controller simulate full physical force and some even simulate musculoskeletal system.

Anonymous

Anonymous Tue 09 Jun 2020 20:17:53 No.11779315 Report

Quoted By: >>11779330

>>11779281
if this is really the case, then you should be able to name which animation controllers you are referring to.

Anonymous

Anonymous Tue 09 Jun 2020 20:22:13 No.11779330 Report

Quoted By:

>>11779315
You already made the point that it takes external forces data to truly make this work.
That's literally all that's missing.

Anonymous

Anonymous Tue 09 Jun 2020 20:24:21 No.11779342 Report

Quoted By:

>>11778978
Reinforcement Learning PhD Candidate here
>>Can this be done? Is anyone working on this at present?
Everything >>11779015 and >>11779100
said is correct. Human kinematic structures vary wildly from standard humanoid robots (i.e. Atlas, Hubo, etc) to the point where learned locomotion has to be self-generated. Unfortunately, simulators suck at modelling friction contacts and RL methods are too sample-inefficient to be brought into the real world, making good old fashioned control theory + planning the best ways forward.

Capcode	All Only User Posts Only Moderator Posts Only Admin Posts Only Developer Posts
Show Posts	All Only With Images Only Without Images
Deleted Posts	All Only Deleted Posts Only Non-Deleted Posts
Ghost Posts	All Only Ghost Posts Only Non-Ghost Posts
Post Type	All Only Sticky Threads Only Opening Posts Only Reply Posts
Results	All Grouped By Threads
Order	Latest Posts First Oldest Posts First

Your latest searches