#robotics Logs

Apr 25 2020

02:31 AM Miyu is now known as hackkitten
05:13 AM rue_bed: --
10:51 AM veverak: hmm
10:52 AM veverak: anybody tried to solve IK with jacobian method?
10:52 AM veverak: rue_bed maybe?
10:52 AM veverak: oh, rue_mohr ^^
10:53 AM rue_mohr: hmm, no...
10:55 AM veverak: hmm
11:03 AM veverak: fu my english again I suppose
11:16 AM mrdata: these are lecture slides; the missing stuff i guess is the lecturer's voice but most things are clear to me
11:19 AM veverak: I wonder about the last step of iteration of the solver
11:19 AM veverak: you do something like: angles = angles + alpha * angle_diff
11:20 AM veverak: I understand that 'alpha' should be small
11:20 AM veverak: I mean, definely alpha < 1.f
11:20 AM veverak: but not sure how to figure out decent value
11:47 AM mrdata: veverak, test and adjust maybe
11:48 AM mrdata: angles of joints closer to the body will have tighter tolerance likely
11:48 AM veverak: it's "generic" solver
11:48 AM veverak: hard to tell when you are close
11:48 AM veverak: I mean, you can easily eyeball working alpha, the efficient and working is tricky
11:49 AM veverak: and I would not be too suprised if it could change
11:49 AM mrdata: will the robot have touch sensors?
11:49 AM mrdata: then feedback gives you good info\
11:49 AM veverak: not really relevant
11:49 AM veverak: :)
11:50 AM mrdata: how so? if you can measure pressure for instance you know when you are gripping something well enough
11:50 AM veverak: yeah, but that is during execution
11:50 AM veverak: my IK solver is for finding precise angles for tips positions generally, not onyl during the execution
11:50 AM veverak: so the sensor is not relevant here
11:51 AM mrdata: if your tolerance were 1 degree, how much slop does that translate to?
11:52 AM veverak: depends on the robot really
11:56 AM mrdata: so does the answer to your question
11:59 AM veverak: not really
12:01 PM mrdata: there is such a thing as close enough; you can construct a formula that uses as parameters the specifics of the robot: eg length of each segment and the range of motion of the joint
12:02 PM veverak: I allredy did...
12:02 PM veverak: :)
12:02 PM mrdata: so then what is left unknown?
12:02 PM veverak: the alpha constant
12:02 PM mrdata: my question is answered by plugging in the specific values for a given robot
12:02 PM mrdata: what is the meaning of alpha
12:02 PM rue_mohr: what about force control
12:03 PM rue_mohr: "close to force"
12:03 PM rue_mohr: vs "close to position"
12:03 PM veverak: rue_mohr: not really interested about that now
12:04 PM rue_mohr: k, doing my bi-directional force feedback thing broke the way I think about a lot of stuff
12:04 PM mrdata: veverak, why not? doesnt it relate to your question?
12:04 PM veverak: nope
12:04 PM veverak: ... again
12:04 PM veverak: I have working method, that uses inverse of jacobian of the FK
12:04 PM mrdata: then what is the meaning of alpha
12:04 PM veverak: you can use it to sovle the Ik numerically
12:05 PM veverak: so, you have actuall set of 'angles', you use FK to get the position of the 'tip'
12:05 PM veverak: let's say the inverse jacobian is called Ji
12:05 PM mrdata: thats understood but what is the meaning of alpha
12:05 PM veverak: angle_diff = Ji * ( target_tip - actual_tip )
12:05 PM veverak: is the step used in the computation of desired angles
12:06 PM veverak: but you can't use just this, because the Ji is correct only "around" actual_tip
12:06 PM veverak: so the real change of angles is: angles = angles + alpha * angle_diff
12:06 PM veverak: alpha < 1.
12:07 PM mrdata: if you guess at alpha = 0.1 then how do you decide whether that is fine enough?
12:07 PM veverak: that is the unanswered question here which I am trying to solve :)
12:07 PM veverak: "how to choose the alpha?"
12:07 PM veverak: I mean, intuitivelly eyeballing soemthing like 0.1 should make it work
12:07 PM mrdata: guess at 0.1
12:08 PM mrdata: then test and adjust
12:08 PM veverak: as it's 0.1 out of the 'error' of tip positions -> it geat smaller with each iteration
12:08 PM veverak: so I do not suspect the thing will oscillate around solution
12:08 PM veverak: but 0.1 also can be slow
12:09 PM mrdata: how slow it is depends on the period of the feedback loop
12:09 PM veverak: wat, nope
12:09 PM veverak: there is no feedback here :D
12:09 PM mrdata: you know the angle somehow
12:10 PM veverak: yeah, you have some "hint" where to start
12:10 PM veverak: you want to precise it
12:11 PM mrdata: ok, well i'm going to stop helping here because you are rejecting my advice. i think you are dwelling in a theoretical matter insatead of making this practical
12:12 PM Tom_L: :O
12:15 PM veverak: ok than :)
12:17 PM Tom_L: ya
06:09 PM rue_mohr: lets keep it to building real robots and not simulations ok?
06:38 PM veverak: from where did this come from? :)
10:47 PM rue_mohr: this?
10:47 PM rue_mohr: build real bots, lots of them