RL dev-log — experiment journal — claudeDrone Team Docs

The rl-lab agent’s cumulative experiment journal — 24 entries (post-Humaned). Each entry is one experiment, with hypotheses, numbers, verdicts, and a glossary.

Timeline (May 2026)

01-02: Bootstrap rl-lab + first PPO run.
03: Frontier reward.
04: RND (intrinsic curiosity) + rllte.
05: Lawnmower baseline (classical heuristic).
06: Retrain max_steps=3000.
07-08: ⭐ Multi-cell forward action — pilot EXP-7, beat the classical baseline 2.15× on 1000 steps.
09-11: Action distribution / soft mask / hard mask gaming (action-space input-importance tests).
12: Recurrent policy — null.
13: Frontier-observation regression.
14: No-scan input-importance test.
15: DR-extended — data-ceiling hypothesis disconfirmed.
16: N-step GAE — sawtooth disconfirmed.
17: TV reward — partial positive.
18: TV reward λ-sweep — LAMBDA05 chosen.
19: OBS-ABL TF-Luna + servo — “distortion is worse than absence” pattern.
20: Map-aug TV mechanism destroyed — null.
21: ⚠ Sim-to-real noise sweep — vl53_0 (forward) critical, without it −79 pp.
22: Rotate-only aug deep regression — augmentation symmetry matters.
23: Session-crash post-mortem — setsid fix.
24: Ent-coef curve — closed plateau cliff.

Related assets

~/drone_media/{rl,sim/rl}/ — 26 evaluation GIFs and graphs (see tmp/media-inventory.md).
Aggregator: rl-agent-logs