# policy-checkpoints/

Final checkpoints of the Appendix-D flow-matching policy (12.6M params:
ResNet-18 visual stem + proprio stem + 4-block transformer encoder + 4-block
flow-matching action decoder, horizon 16), trained on 585 generated
observation/action samples from 5 TACO sequences (5 train / 1 held-out val).

Files:

- `best.pt`, `last.pt` — `torch.save` dicts with keys `model`, `config`,
  `stats` (NO optimizer states). `best.pt` = best val-loss step (~250).
- `best_ts.pt` — TorchScript export of `sample_action` (B=1, normalizer
  buffers baked in), traced on CUDA; loadable in the Isaac Sim venv without
  timm (`torch.jit.load`). Parity vs eager: 6e-8.
- `stats.npz`, `dataset_meta.json` — dataset normalization stats and metadata
  required to run the policy.

Produced by (repo https://github.com/physercoe/egoengine-repro, see
`docs/POLICY.md`):

```
.venv-policy/bin/python scripts/train_policy.py            # bounded run
.venv-policy/bin/python scripts/export_policy.py           # -> best_ts.pt
```

Bounded closed-loop sim eval (`scripts/eval_policy.py`, 5 rollouts/seq):
train seq 4/5 (0.80), held-out seq 0/5 (0.00) — a pipeline proof at 585
samples, NOT a policy-quality result (thin-data caveat, `docs/POLICY.md` §2).

License: MIT (our weights/code). TACO-derived training data: CC BY 4.0.