Skip to content

Introduction

WorldForge is a Python integration layer for testable physical-AI world-model workflows. It gives providers, score models, embodied policies, and media generators explicit capability contracts, then adds world state, planning, evaluation, benchmarking, diagnostics, and CLI tools for local experiments and adapter development.

The project is built around a strict provider boundary:

  • predictive models roll state forward through predict
  • score models rank candidate action sequences through score
  • embodied policies propose action chunks through policy
  • video and media systems expose generate or transfer
  • auxiliary models expose reason or embed only when those operations are implemented directly

This keeps provider semantics honest. A JEPA cost model is not treated as a video generator. A VLA robot policy is not treated as a predictive dynamics model. A media generation API is not treated as proof of controllable physical planning.

WorldForge is for Python developers building provider adapters, local physical-AI experiments, world-model planning loops, evaluation harnesses, and testable prototypes. It is not a hosted control plane and it does not own checkpoints, robot runtimes, production telemetry, or durable multi-writer persistence.

Documentation Map

Need Start here
Install, create a world, and run the mock provider Quick Start
Find an exact CLI command or optional runtime smoke CLI Reference
Understand capability names and provider boundaries World Model Taxonomy
See module responsibilities and planning pipelines Architecture
Record events, worlds, plans, and benchmarks in Rerun Rerun Integration
Validate a checkout, diagnose providers, run smokes, or prepare a release User And Operator Playbooks
Add or promote a provider adapter Provider Authoring Guide
Check package, testing, typing, and ML-runtime quality rules Engineering Quality

If a page changes public behavior, update the linked operational page and the changelog in the same branch. WorldForge docs are treated as part of the package contract, not as post-release notes.