waremax

Glossary

The vocabulary of RMFS simulation, discrete-event modeling, and the WareMax RL benchmark.

RMFS
Robotic Mobile Fulfillment System — a warehouse where autonomous mobile robots carry storage pods to stationary human pickers, in the style of Kiva / Amazon Robotics. WareMax simulates the pod-to-person dispatching layer of an RMFS.
AMR
Autonomous Mobile Robot. In WareMax, robots are AMRs moving on a graph topology with a configurable maximum speed; the simulator does not model AS/RS cranes or AGV tuggers.
Pod-to-person
A fulfillment pattern where robots bring inventory pods to a fixed pick station, rather than a picker walking to the inventory. WareMax models pod-to-person only.
DES
Discrete-Event Simulation — a modeling approach where the simulation clock advances only at discrete events, never continuously. WareMax uses a single-threaded event-queue DES kernel.
Dispatching policy
The rule that decides which robot handles which pick task (task allocation). WareMax ships five heuristic baselines and supports learned policies via the RL interface.
Gymnasium
The maintained successor to OpenAI Gym — a standard API for reinforcement-learning environments. WareMax exposes WaremaxAllocEnv as a Gymnasium environment via PyO3.
Reward mode
How the RL environment scores a decision. WareMax provides four: sparse, dense, attribution, and routed. The routed mode charges only the controllable cost.
Causal delay attribution
A decomposition of each completed task’s cycle time into five buckets — assignment, travel, queue, congestion, service — that sum to the whole. Used for analysis and as an RL reward signal.
ChaCha8 seed
The random number generator (ChaCha8) seeded from a u64 that drives all randomness in the simulator. Fixing the seed is what makes a run reproducible.
Canonical tie-breaking
Breaking ties by entity id rather than by iteration order, applied everywhere in WareMax so that two runs with the same seed produce byte-identical trajectories.
Congestion-aware routing
Routing that accounts for node and edge capacities and a configurable congestion weight, so robots contend for shared graph resources realistically (waremax-map).
SMDP
Semi-Markov Decision Process — the framing WareMax uses for the RL problem, where the agent acts only at decision points (a robot becomes free and a task waits), not on a fixed timestep.
Byte-identical trajectory
The property that the exact sequence of simulator states and events is bit-for-bit reproducible given the same seed and action sequence. WareMax enforces this with tests.