train-policy git:20261010.7768819A

Immutable. This exact content never changes and is served at /api/v1/blob/fa9ddd54c70bb17c.

---
name: train-policy
description: Use when planning, reviewing, or operating robot policy training across LeRobot, Isaac Lab, SONIC, and workflow YAMLs.
---

# Train Policy

## When To Use

Use this skill when a task asks how to train, fine-tune, evaluate, or export a
robot policy through NPA workbench tools. It is the workflow-level entry point
before choosing LeRobot, Isaac Lab, SONIC, or GR00T-specific skills.

## Procedure

1. Identify the policy family and data contract:
   LeRobotDataset for LeRobot, Isaac Lab task config for RL, retargeted motion
   artifacts for SONIC, or model-specific inputs for GR00T.
2. Select the GPU target with `skills/atomic/gpu-selection/SKILL.md`.
3. Configure input and output S3 prefixes. Checkpoints and evaluation artifacts
   must be run-scoped.
4. Choose the executable path:
   direct CLI for a single tool, SDK for application code, or SkyPilot YAML for
   composed training workflows.
5. Verify command help and YAML parsing locally before live GPU submission.

## Three-Tier Contract

- CLI: `npa workbench lerobot train`, `npa workbench isaac-lab train`,
  `npa workbench sonic train`, and related `eval`, `export`, `serve`, or
  `infer` commands.
- SDK: use the workbench SDK modules for application code and shared helper
  functions for request construction.
- YAML: `isaac-lab-rl-train.yaml`, `sonic-train-standalone.yaml`, and sim-to-real
  workflow YAMLs are executable references. The parallel sweep is now the
  `npa.workflow` spec `workflows/testing/isaac-lab-rl-sweep.yaml` (`--runtime`); its
  raw template is retired. GR00T N1.7 training uses the real
  `workflows/main/groot-1-7-finetune.yaml` toolRef path, with `gpu_count`
  propagated into both H100 resources and the upstream trainer world size.

## Gotchas

- Do not route RT-core-dependent training or render validation to H100/H200.
- Do not substitute repository-local output directories for S3 artifact paths in
  public examples.
- Treat tiny smoke trainers as verification substitutes only when the prompt
  explicitly allows minimal production-input substitution.
- Keep W&B, Hugging Face, NGC, and S3 credentials redacted.

## Verify

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

The smoke test invokes training command help and parses the referenced training
YAMLs.

## Public training and deployment recipe

Use `workflows/testing/robot-policy-train-and-serve.yaml` and
`docs/workbench/cookbooks/robot-policy-train-and-serve.md` for real FiftyOne selection,
continued SmolVLA training, two measured gates, portable export, independent GPU
HTTP serving and native LIBERO simulation, then same-run standalone HTML/MP4.
The standard Workbench runtime owns submission, artifact lineage and recovery.
This recipe uses managed Kubernetes and torchrun. The separate Slurm recipe
requires operator training scripts. This public baseline does not prove in-house
foundation pretraining.