sim-to-real git:20261002.bccd515A

Immutable. This exact content never changes and is served at /api/v1/blob/553888507a0d6fbc.

---
name: sim-to-real
description: Use when designing, reviewing, or operating sim-to-real workflows that move data through simulation, policy training, synthetic generation, evaluation, and loop control.
---

# Sim-To-Real

## When To Use

Use this skill for robotics teams that need configurable data and artifact flow
through simulation, policy training, synthetic data generation, evaluation, and
iteration without customer-specific names or infrastructure baked into source.

## Procedure

1. Define a run ID and run-scoped S3 prefixes before launching.
2. Import source robot, scene, or task data into the run prefix.
3. Generate or augment simulation data with configured workbench tools.
4. Train or fine-tune the policy against the run-scoped dataset and write
   checkpoints to S3.
5. Evaluate the policy with deterministic metrics or a configured VLM backend.
6. Decide whether to stop, continue, or route artifacts for review based on
   configurable thresholds.

## Three-Tier Contract

- CLI: use `npa workbench workflow` and tool commands
  such as Genesis, LeRobot, SONIC, MJLab, Retargeting, LanceDB, Cosmos, and
  VLM-eval.
- SDK: keep workflow submission and config materialization in shared helpers so
  notebooks and services use the same artifact paths.
- Workflows: `workflows/testing/vlm-eval-loop.yaml` is the executable reference for the
  VLM-eval gating loop (scores a whole rollout set and writes
  `task_success_report.json`). The staged engine's single maintained end-to-end
  YAML is `workflows/main/sim2real.yaml`; similarly named
  npa.workflow files are explicitly demo-only DSL fixtures.
  `sim-to-real-pipeline.yaml` and `sim-to-real-trigger.yaml` are retired — the first ran
  `npa.workflows.sim_to_real real-loop`, which raises a DeprecationWarning pointing here.

## Field failure navigation improvement

Use `workflows/testing/field-failure-policy-improvement.yaml` for the standard
six-stage capture validation, reconstruction, policy fine-tuning, paired
held-out evaluation, and recommendation flow. Follow
`docs/workbench/cookbooks/field-failure-policy-improvement.md` for the sealed
bundle and adapter contracts. Operators supply real navigation adapters in
immutable BYOF images, captures, baseline checkpoint, and a disjoint held-out
cohort; missing inputs fail closed. GPU qualification remains unverified.

Preserve the same evaluator callable, source digest, image, scenarios, seeds,
and protocol for both policies. Conditional S3 claims isolate attempts;
completed stages revalidate retained evidence before reuse. A failed attempt
requires a fresh run ID. Raw trajectory hashes must be unique within and across
policy arms, including byte-identical deterministic rollouts. Promotion requires
positive primary improvement and no regression beyond the sealed per-episode
thresholds. The result is a recommendation and never authorizes deployment.
Hashes verify identity and consistency, not physical measurement truth.

## Gotchas

- Do not hardcode customer names, event names, personal names, tenant IDs,
  registry IDs, bucket names, VM IPs, or private endpoints.
- Prefer dry-run plans for data movement, autoscaling, and external service
  calls when validating workflow shape.
- Keep artifacts partitioned by run ID so repeated experiments do not overwrite
  each other.
- Use existing workbench tools for S3 sync, model inference, training, and VLM
  evaluation instead of one-off scripts.

## Verify

```bash
npa/.venv/bin/python -m pytest npa/tests/guardrails/test_skills_index.py -q
```

The smoke test parses the sim-to-real YAMLs and confirms the workflow CLI help
loads through the installed package.