search

pantor / ruckig

Motion Generation for Robots and Machines. Real-time. Jerk-constrained. Time-optimal.
https://ruckig.com
MIT License
730 stars 165 forks source link

question about motion planner #85

Closed grotius-cnc closed 3 years ago

grotius-cnc commented 3 years ago

Hi,

First i want to thank the writers and contributers of Ruckig. I used this library intensively for my projects, and i am happy with it. Here are some examples: https://github.com/grotius-cnc/hal-core And a video of a ruckig realtime implementation : https://user-images.githubusercontent.com/44880102/127900012-b7c19051-7d22-45dc-b7ee-0a229e1caa60.mp4

I have a gcode program that is interpolated by ruckig, an sich this works ok. See picture for example, but there is a implementation problem.

In some cases the machine has to slow down at sharp corners, at P4 and P5 the machine has to slow down to avoid a shooting over. At P2 there seems to be no problem. The conclusion, when paths are in line, the machine can stay at 100% speed. If the machine has to make corners, the machine has to slow down.

Can ruckig solve this problem?

My idea was to sample each motion line (in a sort of look ahead function) until there is a minimal (or acceptable) overshoot (Ferror).

I was wondering if anyone else had to deal with this problem. And maybe has a better idea?

Thanks.

problem

pantor commented 3 years ago

Hi @grotius-cnc,

the Ruckig Community Version is not really suited for path following. For a general case, finding velocities and accelerations given a set of intermediate positions is a difficult and complex task. There is/was also a discussion at #65 that might be of interest here.

If you are talking about sharp corners, a jerk-limited trajectory will have to slow down to zero velocity anyway. If some path deviation is ok, your approach would work fine, so this is definitely a way to go in your specific case.

grotius-cnc commented 2 years ago

Thank you for your comment. I will work on this issue coming day's. I have been delayed to solve this.

If some path deviation is ok, The strategy of servo controllers is to follow the given coordinates as good as possible. A difference of encoder feedback and position command is always the case during a run, so far as i know.

Currently i have 2 idea's.

  1. Creating a chunk of c++ code and try to implement senario's with some kind of look ahead function. This has downsides at runtime. It eats memory and processor pieces, and i am already at ~100% load. I also think this calculation has to be done before runtime. So all the optimal "velocity end's" are known.
  2. Creating a Ai fann module and train the machine to optimal results. (difficult course) At runtime, it has the training set with a few values. I think processor is happy. something like this, but then complexer : https://github.com/grotius-cnc/hal_artificial_neural_network_xor2

It looks crazy, but what you see i just one Dofs performing 2 machines, in total 12 axis.

https://user-images.githubusercontent.com/44880102/144765887-f340a338-94d0-4e9b-b308-38a88b8f102a.mp4

pantor commented 2 years ago

Looking forward to results from the FANN module!