Reproducible RL research on warehouse dispatching
WareMax gives RL researchers a deterministic, Gymnasium-compatible warehouse-dispatching benchmark with honest baselines and byte-identical replay, so published results are reproducible.
Warehouse-dispatching RL results are notoriously hard to replicate: non-deterministic simulators, unpublished baselines, and single-seed numbers. WareMax is built to remove those excuses.
What you get
- A Gymnasium env (
WaremaxAllocEnv) with a Dict observation, action mask, and SMDP framing. - Byte-identical replay so a reported trajectory can be regenerated exactly.
- Five heuristic baselines shipped in-box, so a learned policy is measured against something honest.
- Multi-seed sweeps with confidence intervals.
A finding, not a sales pitch
On the built-in presets, trained dispatchers match nearest-robot and round-robin but do not surpass them — the system is capacity- and destination-contention-bound, so state-blind round-robin is near-optimal. WareMax exists in part to make that result reproducible, and to give you the tunable structure to find the regimes where dispatching choice has leverage.
The multi-seed results referenced in the accompanying research live under
crates/waremax-gym/python/results/ in the repository.
Related: comparing dispatching policies · Gymnasium RL loop. Built by Skelf Research.
Frequently Asked Questions
What makes WareMax suitable for research reproducibility?
Every run is byte-identical under a fixed seed and action sequence, the RL loop survives a worker thread via a crossbeam handshake, and five heuristic baselines ship in-box so comparisons are honest.
Does the RL agent beat the heuristics?
On the built-in presets, no — learned dispatching matches nearest-robot and round-robin without surpassing them, because the system is contention-bound. WareMax publishes this as a finding and lets you search for regimes where learning helps.