Bring an existing robotics repository
Use a .him overlay without replacing the repository, simulator, trainer, or agent.
The overlay is the integration
A HIM Package is an immutable envelope around work you already have. It declares identity, entrypoints, runtime/process contracts, artifacts, tests, compatibility, security, provenance, licenses, and publication policy. It does not force a HIM training framework or source layout.
your-robotics-repo/
├── existing source, configs, trainer, simulator, datasets...
└── .him/
├── him.yaml
├── contracts/
├── adapters/
├── samples/
└── tests/Package workflow
# Use only when you explicitly need the advanced package workflow.
him init .
# Simulator repositories: him init . --simulator
# Point the generated entrypoint at the command the project already uses.
him doctor . --json
him validate --json
him test --local --json
# Register the package without launching compute:
him push --visibility unlisted --yes --json- Repeated builds produce identical lock and archive bytes.
- The CLI shows the exact package and committed-source upload before a run.
- Credential-like filenames, escaping symlinks, missing artifacts, invalid contracts, and publication-right gaps fail closed.
- Source is a separate private, digest-bound Git snapshot. Weights, datasets, endpoint addresses, and calibration stay in declared artifacts behind object-level authorization.
- Most projects upload all committed source outside .him. Large multi-project repositories can set extensions.source_paths to an exact list of committed files and directories so unrelated workspaces are not uploaded.
- Hosted commands receive HIM_CANDIDATE_OUTPUT, HIM_EVALUATION_OUTPUT, and HIM_MEDIA_OUTPUT. Existing artifacts/result.json outputs are preserved automatically for compatibility.
extensions:
source_paths:
- robotics/assets
- robotics/train
- motions/dance.csvEntrypoints preserve native work
| Purpose | Examples | HIM expectation |
|---|---|---|
| Inference | Python module, OCI command, policy service, private endpoint | Typed observations/actions or declared batch inputs/outputs |
| Evaluation | MuJoCo episode driver, Isaac play task, regression harness | Frozen candidate output and isolated evaluator contract |
| Training | torchrun, JAX/Orbax, native lab pipeline, multi-stage DAG | Checkpoint/resume, budgets, promoted candidate outputs |
| Render | Simulator camera or trajectory renderer | Factual media derived from recorded state |
| Deployment | Booster/Unitree adapter bundle | Separate hardware readiness and safety gate |
Use the same contract with coding agents
Use Copy for AI at the top of any guide to generate a complete setup prompt for Codex, Claude Code, Cursor, or another coding agent. The prompt tells the agent to inspect first, preserve native commands, edit only .him by default, validate locally, and ask before it uploads source or spends credits.