⚠ See the matched-control correction (Part 49) — the matched-control result revises this: the “beats PPO” claim here was vs an under-budgeted PPO; against a budget-matched PPO the abstraction wins sample-efficiency only, not the asymptotic ceiling (OpenCabinet stays a ~7× sample-efficiency win at a tie on success).

Part 40: the rigid structured abstraction caps at 0.77 on OpenCabinet (PPO 0.98). Part 41: warm-start then release uncaps it to PPO parity, but the 62M bootstrap meant no sample-efficiency win. Part 42 closes it: make the bootstrap lean, and it’s a clean beat.

The lean curriculum

  • Bootstrap: restore from the 6.5M skillD checkpoint — the leanest point where the structured skill already reaches the handle (~0.98 reached). (No retrain: the skillD checkpoints already exist at every 3.3M.)
  • Release: continue as a free end-to-end policy (action = clip(policy(obs)), no hand-offs), LR = 5e-4 (lowered from 1e-3) as a stabilizer against the Part-41 late-collapse.

Result (true success, env metric box_target ≥ 0.9, n=256, sustained over last 3 checkpoints)

seed peak success total steps to sustained ≥0.95
s1 1.000 6.5M + 4.9M = 11.4M
s2 0.984 6.5M + 1.6M = 8.1M
s3 0.981 6.5M + 3.3M = 9.8M
s4 0.981 6.5M + 8.2M = 14.7M
— PPO 0.9883 @ 29.5M

All 4 seeds reach 0.98–1.0 in ~8–15M total env-steps (mean ~11M) vs PPO’s 29.5M — a ~2.5× sample-efficiency beat at matched/better success. The LR=5e-4 change fixed the late-collapse: all 4 seeds now sustain (vs 1/4 collapsing in Part 41).

Why this is the real beat

It satisfies both axes honestly: success parity (0.98–1.0 ≥ PPO 0.9883, s1=1.0) and a total-step win (bootstrap + release counted in full: ~11M ≪ 29.5M). The structured abstraction is the accelerant — a 6.5M bootstrap that gets reach + a coarse skill — and the release removes the rigidity that capped the ceiling (Part 40). Bootstrap fast, then release: the abstraction’s structure buys the sample-efficiency, and the free late phase buys the ceiling.

This extends the campaign’s law to the hard task: on PickCube the in-loop abstraction ties PPO and beats it on sample-efficiency (prior fits); on OpenCabinet the curriculum form now does the same (beats PPO 2.5× on sample-efficiency at matched success), where the rigid structured form capped (Part 40) and the matched form broke even (Part 41).

Caveats

4 seeds; total steps = 6.5M bootstrap + release-steps-to-sustained-0.95 (fair; the bootstrap reuses an existing skillD checkpoint but the 6.5M is counted in full vs PPO from-scratch). Success = same env metric/scorer as the PPO baseline (n=256, 1000-step), read from release/EVAL_boot6p5_s*.json (never shaped reward), sustained over the last 3 checkpoints. A complementary result: a TAMP controller (motion-planned reach + IK-push) reaches ~0.83 success analytically (no learning) — Part 43. Prior: Part 40 (cap), 41 (uncap/match), 27 (PickCube beat).