Warehouse Fleet Simulator
Command a fleet of warehouse robots hauling live orders — flip between a central dispatcher and decentralized swarm negotiation and compare the traffic.
How to use
Orders stream in on their own: robots claim them, pull cartons from the racks, and run them to the packing stations on the right wall, topping up at the left-wall chargers when their battery runs low. Flip the Coordination switch to compare a decentralized swarm against a central dispatcher under the same load, push the sliders until the order queue fills, and click any robot to fail it and watch the fleet absorb the loss. In first person, Take control hands you the wheel — park across a busy aisle and be the disturbance the swarm has to solve.
What you are watching
The core question of multi-robot coordination: who decides? A central dispatcher with global state makes near-optimal assignments and smoother traffic, but it reacts on a fixed cycle and is a single point of failure. The swarm inverts that: every robot bids for work with local information only, assignments are merely good, aisles get jostled — and the response to any disturbance is immediate, local, and nobody’s job. The dispatcher usually wins on congestion; the swarm wins on recovery. That tension — optimality versus resilience — is the real engineering decision behind every warehouse fleet.
Everything on this floor is a named idea from the multi-robot literature. Job assignment in swarm mode is a single-item auction — the market-based answer to what Gerkey and Mataric’s taxonomy classifies as single-task, single-robot, instantaneous assignment — while the dispatcher plays greedy stand-in for the optimal assignment problem that the Hungarian and auction algorithms solve exactly. Every route is plain A* over the aisle grid, Hart, Nilsson, and Raphael’s 1968 algorithm doing what it has done ever since. The traffic layer — claim the next cell, yield, replan around a jam — is a lightweight cousin of Silver’s windowed cooperative A*, and the general problem it deliberately avoids solving optimally is multi-agent path finding (MAPF). The floor itself is the architecture Wurman, D’Andrea, and Mountz described for Kiva’s systems — the paper that moved this whole genre from research to industry. The references below are the right doors in.