A packing tree that teaches RL where the boxes go

What it does

This is the reference implementation of “Learning Efficient Online 3D Bin Packing on Packing Configuration Trees” (ICLR 2022). Items arrive one at a time, and a deep reinforcement learning agent must place each into a 3D container immediately — no peeking at the future. The trick is representing the bin’s state as a hierarchical packing configuration tree (PCT), which the DRL policy traverses to pick a placement. Compared to the authors’ earlier DRL packing work, this version supports arbitrary container and item sizes, a continuous solution space, and approximate packing-stability checks.

The interesting bit

The tree structure is what makes practical constraints tractable: instead of the policy choosing raw coordinates in a continuous space, it walks a hierarchy of candidate configurations, which keeps the action space sane and training stable. The repo also ships heuristic baselines (LSAH, OnlineBPH) with sources cited in the code — a courtesy the field rarely extends.

Key highlights

• Continuous-domain online 3D BPP environment, plus the discrete one
• Approximate stability algorithms (linked to the authors’ separate work on arXiv)
• Pretrained models and a 3000-trajectory dataset available for download
• Training data generated on the fly; tensorboard logs included
• Part of a small family of repos (DRL, PCT, IR-BPP, Packing-Tools) covering lookahead, buffers, irregular shapes, and rendering

Caveats

• Pinned to an old stack — the authors suggest Python 3.7, gym 0.13, torch 1.10 — so expect dependency friction on modern environments
• Pretrained models only cover the discrete setting with a (10,10,10) bin; continuous-domain users must train their own
• Datasets and models live on Google Drive links, not in the repo

Verdict

Worth a look if you work on packing, logistics, or RL for combinatorial problems and want a well-cited, actively-maintained-family baseline. If you just need a quick heuristic packer, the included heuristic.py baselines may be all you need.