WorldFuzz

Run a bounded robotics task, keep its evidence, and reuse checked results or data. WorldFuzz connects a local workspace to the same platform service through a browser, Python SDK, CLI and MCP. A run keeps exact inputs and versions, rights, resource limits, task outcomes, verification and downloadable artifacts together.

Free preview: 0.4.0rc6, Python 3.11+. Download the original archives from GitHub Releases and read the static documentation. PyPI and MCP Registry remain unpublished. See PUBLIC_PREVIEW for the selected real NVIDIA/Nebius navigation result and publication scope. WP33 remains ACCEPTED_TESTED_SCOPE for the retained rc1 native and rc2 Kali container receipts; later candidates do not repeat that matrix. See the release scope for changes and retained evidence limits.

Start with the platform

Download the wheel and SHA256SUMS from the release above and verify the file hash. In a new Python environment, install the supplied wheel. The package resolves its pinned base dependencies; installation needs network access unless they are cached:

python -m pip install ./worldfuzz-0.4.0rc6-py3-none-any.whl
python -m worldfuzz platform --root workspace/results --store workspace/catalog.sqlite3 serve

Use a new workspace/ directory. Open http://127.0.0.1:8766/ for the UI and http://127.0.0.1:8766/docs/ for the English documentation. No frontend build, model key, GPU, device or HDF5 download is needed for published grid navigation. The server stays on loopback. Organization workspaces require a scoped login.

For one new task, open a second terminal in the same directory/environment:

python -m worldfuzz platform --root workspace/results --store workspace/catalog.sqlite3 run navigation --resource movingai:0 --case-index 0 --key quickstart-nav-0 --max-seconds 30 --cpu-seconds 20 --memory-mib 256 --artifact-mib 1 --wait-seconds 30 --export workspace/export-nav-0

This uses run_goal and the shared runner. It checks a route against original Moving AI inputs and the independent route checker, then returns a run ID and separate run/task/verification outcomes. The new export directory contains the summary, JUnit, result and versioned JSON artifact views. Original artifact bytes remain available from the authenticated service. Reuse the same goal/key after a wait timeout; do not create a second job or overwrite a completed export. Stop the server you started with Ctrl+C when finished.

Related MCP server: ai2robot-mcp

Choose a supported process

Integration and workspace guide explains client choices. Connect an MCP client and install the optional Agent Skill for one bounded task with the existing Python stdio server. The served Supported tasks pages give each exact adapter operation, profile, requirements, bounded example and evidence scope. API and Profile reference are generated from the existing contracts, not a second manual schema copy.

Read, compare and export

The UI separates a new run from a saved recording. compare_results compares compatible completed results; it does not run another task. export_result prepares a plan only. Export is confirmed by completed execution and existing hash-checked files. Source rights remain separate from workspace membership; linked pairs and held-out splits remain protected.

Develop an installed extension with the adapter guide and contribution process. The separately supplied developer kit works without a source checkout. Compatibility, support and local usage metrics state the current limits; no external-adoption or telemetry claim is implied.

See historical WP33 installation and recovery evidence, organization access, workspace recovery and documentation export instructions. Documentation is ordinary HTML without mandatory JavaScript and can be exported with a configured base URL. Preparing a sitemap or llms.txt is not publication or indexing.

Evidence and status

M7 is ACCEPTED_LIMITED_SCOPE. The twofold action-reduction target was not achieved; both independent agent studies retain their original failed verdicts. Software integrity, offline metrics and simulator replay do not prove physical robot behavior or universal safety.

• Release scope, historical acceptance and limits

• Historical dashboard and experiment commands

• Saved published navigation report

Historical reports are recordings, not fresh execution. The old dashboard command remains available through the historical guide; it is separate from the current platform workflow. MIT license; third-party source notices remain with their inputs and the delivered package.