DishwasherBench is a playable, browser-accessible dishwasher-loading game called Spotless built on a deterministic physical simulation of a real Bosch SMV4HAX40G. Players or agents place 3D models of dishes into the machine and run a simulated Eco 50 wash that includes jets, droplets, run-off, soils and detergent chemistry. The system awards points for items that emerge clean and includes seven difficulty levels, a catalogue of items and soils, deterministic rigid-body loading, and both a slow official tier (accurate scoring) and a faster preview tier. Runs are reproducible episode records that can be replayed; the official wash of a largest level takes about 4.9 minutes on the referenced MacBook Pro. The repository provides three play modes - web page, MCP server for agents, and a harness for scripted players - and two models have already been tested via an OpenAI-compatible endpoint.
Validation is explicit but partial: the simulation was calibrated on one test wash and tuned to published class means of a sister machine; 48 literature-derived loading phenomena were checked (17 pass, 21 fail, 9 unsupported by current mechanisms). Water is conserved exactly and the fast tier agrees with the official tier on most verdicts, but scoring is reliable today only for orientation-sensitive items (glasses, mugs, pots) and not yet for plate and cutlery stacking. Missing or unmodeled aspects include drying, glass film, items moved by jets, and broader calibration beyond Eco 50. The project is an early 0.1.0 release with specific system requirements (Node 20.19, pnpm 9, WebGPU-capable GPU, Chrome) and is intended as a benchmark platform for language-model agents with known limitations.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.