DishwasherBench – Can Frontier Models Stack a Dishwasher?
Summary
DishwasherBench is an open-source dishwasher-loading game built to benchmark language models. It simulates loading a real Bosch dishwasher with rigid-body physics and scores items after a wash. The project provides three ways to play: in a browser, via an MCP server for agents, and a scripted benchmark harness; it includes calibration, validation against real washes, and a test suite of phenomena. The repo documents requirements, quick-start steps, and model results for OpenAI-compatible endpoints, enabling reproducible AI testing on an open-source platform.