Factory Bench (MRTA + MAPF)
A test bench for a heterogeneous multi-robot factory (semester project, Prof. Pranjal Vyas).
- MRTA (task allocation + scheduling): decides which robot carries which product, in which order.
- MAPF (multi-agent path finding): plans collision-free, time-stamped
(x, y, t) routes on the factory map.
- The bench runs both together, feeds path delays back into the schedule, and replans when a robot fails or an aisle is blocked.
- Algorithms are swappable: built-ins, or your own Python file uploaded in the browser.
One app, two deliverables: the MRTA window (/mrta) and the MAPF window (/mapf) each show one module on its own; the main page shows both together.
Status (honest)
- Working and tested: engine (33 tests), 4 MRTA algorithms (greedy, auction, genetic, tabu), 3 MAPF algorithms (independent baseline, prioritized, CBS with fallback), events (blocked cells, robot failure/recovery), validators, in-browser Python (Pyodide), Run/Import/Compare tabs, module windows.
- The sample factory (
scenarios/small_factory.json) uses made-up numbers. It is not from the professor.
- The Gemini import path (
/api/convert) is written but not yet tested against the real API (no key during development). Without a key it answers 501 and the deterministic importers (.py, .json, ASCII grid) still work.
- The Sambhav prototype was never available, so the MRTA side is a fresh implementation of the same idea.
- Not decided yet: official metrics, the meaning of "Direct/Indirect" in the notes, final deliverable wording.
Layout
python/bench/core grid, scenario, tasks, schedule, validators, orchestrator, plug-in loader
python/bench/mrta greedy, auction, genetic, tabu
python/bench/mapf lowlevel (space-time A*), independent, prioritized, cbs
python/tests pytest suite
scenarios/ sample factory + events
plugins-examples/ templates to copy for your own MRTA / MAPF algorithm
web/ Next.js app (UI, Gemini route, Pyodide worker)
Run locally
cd python && python3 -m pip install pytest && python3 -m pytest -q # engine tests
cd ../web && npm install && cp .env.example .env.local
npm run dev # http://localhost:3000
npm run dev/build first copies the Python engine and the Pyodide runtime into web/public/ (git-ignored), so the browser loads everything from your own domain.
Environment variables (set in Vercel, never in the repo)
| name | purpose |
|---|
GEMINI_API_KEY | your Gemini key (server only) |
GEMINI_MODEL | model id you want to use, e.g. the one shown in Google AI Studio |
APP_ACCESS_CODE | optional; if set, converting uploads needs this code |
Deploy on Vercel
- Vercel dashboard > Add New > Project > import this GitHub repo.
- Root Directory:
web. Framework: Next.js (auto-detected).
- Add the environment variables above, then Deploy.
Your own algorithm
Copy a file from plugins-examples/, edit it, upload it in Import. One file = an INFO dict + allocate(...) (MRTA) or plan(...) (MAPF). Uploaded code runs only in a Web Worker in the visitor's browser with network access blocked by the page's CSP, under a soft time limit and a hard stop button.
Model assumptions (to confirm)
Whole time steps; a robot with step cost k needs k steps per cell; conflicts = same cell at the same time, or swapping cells (following is allowed); a robot with no work left leaves the floor; a failed robot stays as an obstacle.
