cogame-rware-warehouse

cogame-rware-warehouse

A port of semitable/robotic-warehouse (RWARE) to a Coworld, with four ROBOT DRIVERS instead of four per-tick RL policies.

Four robots share a small warehouse. Shelves stand in blocks; the aisles between them are one cell wide. Two workstations sit at the bottom of the middle lane. A station board shows the requested shelves, and a robot must drive to a requested shelf, drive under it, lift it, carry it to a workstation, and then carry it back to an empty storage slot and put it down — because a loaded robot cannot pass under another standing shelf and cannot pick anything else up until it has stowed what it holds.

Nobody scores alone. The number the league reads is the number of shelves the fleet delivered, and the only way to lose it is to jam the aisles.

A policy is just a prompt

A seat is an LLM if it sets PLAYER_PROMPT, and a scripted baseline if it sets PLAYER_SCRIPTED=shuttle|courteous. Both come out of the same image, and the game server — not the player container — makes every model call.

Every 20 ticks each driver is handed its robot's view of the warehouse and answers with one JSON order:

{"verb": "fetch", "shelf": "S12", "say": "taking S12, keep column 7 clear",
 "notes": "then stow at (2,7)"}

fetch · deliver · stow · yield · hold — five orders, and a deterministic pilot drives the robot until the order finishes or the driver changes it. say is a fleet radio call every other robot hears next turn; it is the only channel through which intentions travel, and it is what makes ad-hoc coordination in a one-wide corridor a real problem.

What makes it hard

You drive one robot and you cannot control the other three. Two robots that meet head-on in a one-wide aisle both stay put — forever — until one of them yields. A line of robots can move together if the one in front has somewhere to go. The workstation queue lane is where every delivery jams, and a fleet that deadlocks there plays out its 500 ticks with the jam flag lit and scores zero.

Scoring

delivered[s]   = requested shelves delivered by seat s's robot
teamDelivered  = sum over s of delivered[s]
scores[s]      = 100 * teamDelivered + delivered[s]

Higher is better; no term is ever negative. The first term is identical for all four seats — pure common interest. The second exists only so the ladder is not a draw machine, and it is deliberately an epsilon: a full round trip takes at least 12 ticks, so delivered[s] < 100 always and the ordering is strictly lexicographic — team throughput first, own deliveries only as a tie-break.

results.win[s] is teamDelivered >= parDeliveries, the same boolean for all four seats, and results.winner is always null: a cooperative episode has no winner.

Two name spaces

In-game the seats are Alpha, Bravo, Charlie and Delta. Those aliases are the only names that appear in an observation, a prompt, an order, a radio line or a sprite label, so a driver can never learn who it is working with. The seats' real policy names live only in results.names, in the replay's join records and in the viewer's scorebug.

Watching

Replays are a static wasm bundle, never a pod: the same Nim sim module the server runs is compiled to WebAssembly and re-derives every frame in the browser from the recorded orders, checking a per-tick hash as it goes. The viewer draws the whole warehouse edge to edge with the request rail in the top band, the delivered counter in the clock, a labelled jam chip, a deliveries sparkline with the jam spans shaded behind it, and clickable beat markers for every delivery, jam, fallback and the end.

Layout

Building

nim c -d:release --path:src -o:rware-warehouse src/rware_warehouse.nim
nim c -r --path:src tests/tests.nim          # the whole suite
docker build -t coworld-rware-warehouse:latest .
./tools/build_replay_viewer.sh "$PWD/dist/static-replay-viewer"

MIT licensed. The vendored upstream sources keep their own licence in vendor/LICENSE-rware.