train-motion-policy

name	train-motion-policy
description	Train, fine-tune, resume, evaluate, and export a HoloMotion motion policy. Use for dataset selection, configuration preflight, checkpoint lineage, distributed training, evaluation coverage, export parity, and preparation for simulation or deployment.

Train Motion Policy

Treat training as one contract-preserving pipeline:

validated data -> training or fine-tuning -> evaluation -> export -> simulation -> deployment gates

Discover the current documentation, environment bootstrap, configuration composition, entry points, evaluator, and exporter from the checked-out source. Use holomotion-train-interpreter to run the selected environment.

Choose training or fine-tuning

• Prefer fine-tuning when a compatible released or validated checkpoint exists.
• Use training from scratch only when requested or when the pretrained policy contract is incompatible.
• Verify checkpoint completeness and optimizer/model-state compatibility before fine-tuning or resume.

Inspect the fully composed training config rather than relying on remembered defaults. Confirm:

• every configured dataset exists, is immutable for the run, and has intentional sampling;
• robot, DOF order, observations, actions, history/future reference, and network module agree;
• experiment name and output directory are distinct;
• resume and fine-tune semantics match the user's intent;
• requested environment count fits available memory.

Preflight

Resolve the current launcher from documentation and source. Prefer its non-submitting dry-run or command-print mode; otherwise inspect it and run the smallest local configuration that validates composition and startup. Review the resolved command, checkpoint, config, data overrides, topology, and output location before starting a long run.

Do not edit shared base configs merely to launch one experiment. Prefer a dedicated config or explicit overrides supported by the current configuration composer.

Evaluate and export

Do not judge a policy from training reward alone.

1. Resolve the evaluator and export contract associated with the selected training config rather than assuming a script or flag name.
2. Require complete intended checkpoint, dataset, and clip coverage.
3. Inspect per-clip failures and balanced aggregates, not only a global mean.
4. Use machine-readable reports for exact values and human-readable summaries for review; discover their current schemas from the evaluator.
5. Confirm exported inputs, outputs, state, preprocessing, and numerical behavior against the training implementation.
6. Run the currently supported simulation evaluation before real-robot use.

Keep training, evaluation, and deployment observation/action contracts aligned. If exported behavior differs from PyTorch evaluation, localize the first divergent stage before tuning the policy.

Completion

Report the resolved config, datasets, checkpoint lineage, command, topology, outputs, evaluation coverage, exported artifact, and unperformed gates. Never claim real-robot success from training or simulation evidence.