WorldFuzz documentation

Run bounded robotics tasks, keep evidence and reuse checked data.

WorldFuzz turns a supported robotics task or an existing recording into a bounded run with its inputs, versions, permissions, resource usage and checked artifacts kept together. Use the same service from a browser, Python, a terminal or an MCP agent.

Run one published navigation task Choose a supported process

What you need

• Python 3.11 or newer on a tested Windows or Linux placement; an operator-supplied wheel or an authorized source checkout.
• A writable private workspace and permission to use its source data. Published Moving AI inputs are bundled; Can HDF5 and personal recordings are not.
• Finite time, CPU, memory and artifact limits. Optional simulation and dataset readers need their pinned dependencies.

What you receive

A durable run ID, separate execution and task outcomes, a verification report, source/profile versions, artifact hashes and a usage ledger. You can compare compatible results, download original files and build versioned datasets while preserving rights and linked groups.

Evidence you can interpret

A saved report is a recording. Reading it is not new execution. A new run passes compatibility, owner permission, source rights and resource checks. An integrity check validates stored bytes; a supported metric checker recomputes a defined result. Neither proves physical robot behavior.

Current scope

M7 remains ACCEPTED_LIMITED_SCOPE; its original action-reduction and agent studies retain their negative results. WP33 is ACCEPTED_TESTED_SCOPE for its retained rc1 native checks and rc2 Kali container receipt. This documentation build does not rerun that matrix or claim full agent certification or a twofold speedup.

No public website, PyPI release or MCP registry listing is asserted. A configured canonical URL is publication preparation, not evidence of hosting or indexing.