#robotics Logs

Apr 29 2020

03:15 AM Jak_o_Shadows: Is this a series actuator?
03:36 AM veverak: Jak_o_Shadows: yes, robot's leg
03:36 AM Jak_o_Shadows: ah, cool.
03:36 AM Jak_o_Shadows: Yeah, jacobian seemed the simplest last I looked into things.
03:36 AM Jak_o_Shadows: What about constraints though?
03:37 AM veverak: not resolved for now
03:38 AM veverak: (not there yet, but I have seen some approaches to that)
03:43 AM Jak_o_Shadows: Honestly, I think random restarts might do it.
03:43 AM Jak_o_Shadows: Or rather, just don't let it move past constraints, and then do a few random restarts
03:44 AM Jak_o_Shadows: At like, the other end of the constraint that failed
03:44 AM veverak: based on the .pdf
03:44 AM Jak_o_Shadows: That'd get you like, a 90% olution
03:44 AM veverak: you can write the math so it optimizes not only towards the 'target tip position'
03:44 AM veverak: but other constraints
03:44 AM veverak: and what is done, is that they introduce another constrians: move joints to their 'mid' position
03:45 AM veverak: so they entire thing tries to move to target tip position and away from joint extremes
03:45 AM Jak_o_Shadows: So that's a rather more complex jacobian?
03:45 AM veverak: wait
03:46 AM Jak_o_Shadows: Doesn't the "Jacobian" method basically just do gradient descent?
03:48 AM veverak: yeah
03:48 AM veverak: jacobian is same, the step is different
03:49 AM veverak: it turnsout that (I - J^-1 J) matrix performs a projection onto the nullspace of J
03:49 AM veverak: so J(I - J^-1 J)G = 0, for any G
03:49 AM veverak: so the actual formula is: angle_delta = J^-1 e + (I - J^-1 J) G
03:50 AM veverak: where G somehow represents second goal during the movement
03:50 AM veverak: and link to articles that used that for various purposes
03:50 AM srk: <3
03:50 AM srk: veverak: links? :)
03:50 AM veverak: but I just woke up, so do no want any details from me :D
03:50 AM Jak_o_Shadows: Oh yeah. So kinda gradient descent, but you're minimising multiple things
03:50 AM veverak: srk: sec
03:50 AM veverak: Jak_o_Shadows: yes
03:51 AM veverak: srk: http://scholar.google.cz/scholar_url?url=https://www.academia.edu/download/34612234/iksurvey.pdf&hl=en&sa=X&scisig=AAGBfm3CCovZfM0u7ZwoM63n9z42xtM2zw&nossl=1&oi=scholarr
03:51 AM veverak: is this reachable ?
03:51 AM srk: Oops! It looks like you're in the wrong aisle. :D
03:51 AM srk: found it
03:51 AM veverak: good
03:52 AM srk: ty!
03:57 AM veverak: Jak_o_Shadows: the best thing is that based on the article, this shit should handle arms with a lot of DOF and multiple spearate end effectors
03:58 AM veverak: + I don't really see any reason to stick only to 'rotational joints'
03:59 AM veverak: given that, it should be possible to have "motion execution" coded for quite a big set of different leg configurations (fingers.., high DOF count, non-rot. joints...)
03:59 AM veverak: and my motion planning can work with similar stuff (but I would have to ivnest some time into working with complexity problems.... stuff is n^DOF now :/)
04:08 AM srk: generic! \o/ :D
04:20 AM Jak_o_Shadows: I quite like the idea of non-rotational joints
04:21 AM Jak_o_Shadows: Also yes, link not working
04:29 AM srk: Jak_o_Shadows: https://www.math.ucsd.edu/~sbuss/ResearchWeb/ikmethods/iksurvey.pdf
04:30 AM srk: I was quite confused about what kind of joints consider as well, you either use standard ones or just rotation / translation or something flexible that covers everything
04:31 AM srk: *joints to consider
04:36 AM Jak_o_Shadows: thanks
04:36 AM Jak_o_Shadows: lately I favour genetic algorithms - but I'm not planning on-line planning
04:52 AM zhanx_ is now known as zhanx
09:35 AM rue_mohr: hmm on one hand I want a robot to do some ik for, on the other hand I have 5 robots ready that could take some
09:36 AM rue_mohr: that said I think they all need their controllers recalibrated
09:36 AM zhanx: yes
09:37 AM zhanx: my servo's will be here soon
01:42 PM LarchOye1 is now known as Guest83850
11:21 PM mrdata_ is now known as mrdata