Giters

alexfrom0815 / Online-3D-BPP-PCT

Code implementation of "Learning Efficient Online 3D Bin Packing on Packing Configuration Trees". We propose to enhance the practical applicability of online 3D Bin Packing Problem (BPP) via learning on a hierarchical packing configuration tree which makes the deep reinforcement learning (DRL) model easy to deal with practical constraints and well-performing even with continuous solution space.

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Usage in real-world

xzpk22 opened this issue · comments

commented

Hello ! Thank you very much for this wonderful work!
I'd like to reproduce the experiment with a real robot (C-6 of your paper). I'm having trouble adjusting the code in order to continuously feed the PCT pipeline new boxes.
The random boxes seem to be generated in PackingDiscrete.gen_next_box, but the function is called at the end of PackingDiscrete.step, in PackingDiscrete.cur_observation().

The ideal workflow, as I understand it, would be :

  • Capture a new Box (sx,sy,sz)
  • Call PackingDiscrete.gen_next_box with box dimensions
  • Generate the observation vector and forward the network with it
  • Call env.step() with the selected_leaf_node
  • Get the box target location from the latest box added in env.packed

Do you think that would that work correctly ? Did you do something similar for your real-world packing video ?
If you still have some code of that experiment somewhere, I would be very glad to consult it :)

Thanks again for sharing your work and for your time!

commented

Hello, thank you very much for your attention. I think your understanding is very accurate. When we pack in the real-world, we can modify it like this. We have some projects that deploy packing robots in the real world, which is based on this logic. of. I'd be more than happy to help, unfortunately the code for the robot project cannot be released due to some non-disclosure agreements. This change is correct and easy to implement, so you can try it out.