# rl-checkpoints/

Residual-PPO (rsl_rl) checkpoints and full logs for the EgoEngine refine
chunks that replay/MPC could not fix (`pending_rl` escalation, see
`docs/REFINE.md` §7 in the code repo).

Layout:

- `<triplet>/<seq_id>/chunk_<a>_<b>/` — local run: `logs/model_final.pt`
  (rsl_rl checkpoint), tensorboard event files, `train_log*.txt`,
  `eval_residual.json` + `eval_residual.png`, `chunk_start_state.npz`.
- `a800/<triplet>/<seq_id>/chunk_<a>_<b>_<tag>/` — remote 4xA800 batch runs
  with the RL.md §5 improvement bundle (dense-after-violation, contact-or,
  warm-start-mpc, feasibility curriculum; `v2_m05f15` = mass x0.5 /
  friction 1.5 sweep variant).

Produced by (repo https://github.com/physercoe/egoengine-repro, see
`docs/RL.md` + `docs/A800_RL.md`):

```
OMNI_KIT_ACCEPT_EULA=YES .venv-sim/bin/python scripts/train_residual.py \
  "<triplet>" <seq_id> --chunk <a>:<b> --dense-after-violation --contact-or \
  --warm-start-mpc --curriculum-start-scale 2.0
# batch: scripts/run_rl_a800.sh
```

Honest status: the stage is delivered but under-budget — the canonical
knife-slip chunk `[40,60)` was NOT fixed (tool feasible_frac 0.0 before and
after RL, see `eval_residual.json`); candid negative result, analysis in
`docs/RL.md`.

License: MIT (our checkpoints/logs).