⚠ 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 this is a sample-efficiency win (OpenCabinet ~7× at a tie on success), not an asymptotic-ceiling beat.

The homepage update laid out two routes to beat PPO on a hard manipulation task: (1) abstraction-as-curriculum (Part 42), and (2) TAMP as the principled controller — a planner that is also an abstraction, kept in the loop with a learned residual. Part 43 delivers route 2, and it turns out to be the most stable beat we have.

The method

ResidualTampCabinet: a task-and-motion-planning analytic controller for PandaOpenCabinet — collision-aware reach, a good-branch grasp command applied directly, and an IK-based pull — kept live at every timestep, with a small PPO residual learned on top:

\[a = \mathrm{clip}\big(a_{\text{TAMP}}(s) + \alpha\,\pi_{\text{res}}(s)\big),\qquad \alpha = 1.0\ \text{(persistent)}.\]

The TAMP controller alone solves the drawer task 0.83 of the time (n=256, env box_target ≥ 0.9, 150-step) — a competent prior, not a solution. The residual is trained with stock brax PPO inside the wrapper. No --save_full_state; success read from res_eval_s*.json (env metric, never shaped reward).

Result (n=7 seeds, env success box_target ≥ 0.9, n=256)

seed peak final sustained ≥0.95 (last 3) steps to 0.95
s1 0.981 0.981 ✓ 3.3M
s2 0.981 0.981 ✓ 1.6M
s3 0.981 0.981 ✓ 1.6M
s4 0.981 0.981 ✓ 1.6M
s5 0.981 0.981 ✓ 3.3M
s6 0.981 0.981 ✗ (wobble, no collapse) 3.3M
s7 0.981 0.981 ✓ 1.6M
— PPO 0.988 — 29.5M (first ≥0.95)
— TAMP-alone 0.83 — 0 (analytic)

7/7 seeds reach 0.981; 6/7 sustain ≥0.95; all cross 0.95 by 1.6–3.3M env-steps versus PPO’s 29.5M — a ~9–18× sample-efficiency beat at matched success. (0.9805 = 251/256: ~5 eval-config resets are the shared ceiling both PPO and this method hit.)

Residual-on-TAMP vs curriculum: the stability win

Both routes beat PPO’s sample-efficiency on OpenCabinet, but they differ on robustness:

route seeds sustained ≥0.95 hard collapses (→0) reach 0.95
curriculum (Part 41/42, lean boot) 5/7 1 (a seed collapses to 0) ~8–20M total
residual-on-TAMP (this part) 6/7 0 1.6–3.3M

Residual-on-TAMP is faster and more stable: zero hard collapses across 7 seeds (the curriculum’s released free policy can drift off its bootstrapped optimum; the in-loop TAMP controller never lets the residual fall that far). Keeping a principled controller live — rather than releasing to a free policy — is what buys the stability.

What it means

This is the campaign’s cleanest beat-PPO result on the hard task, and it validates the “TAMP is simultaneously a planner and an abstraction” thesis: a general, legible geometric/symbolic controller is a better in-loop prior than a hand-coded phase machine (Part 40’s structured skill capped at 0.6–0.8; TAMP+residual sustains 0.98). The residual’s job shrinks to polishing a controller that already nearly solves the task — which is exactly why it is sample-efficient and stable. Where the prior fits (PickCube tie + 1.7×, OpenCabinet via curriculum or TAMP), abstraction-in-the-loop beats PPO on efficiency and stability; where it does not (Part 47, AcrobotSwingup), it backfires. The lever, throughout, is prior quality.

Caveats

7 seeds, single protocol (150-step, n=256, env box_target ≥ 0.9), α=1.0 fixed. s6 final eval dipped below the 0.95 last-3 bar but did not collapse (peak/final 0.981). TAMP-alone 0.83 is the good-branch controller (RESULTS.json). PPO bar 0.988 @29.5M is the matched-protocol Part 42 baseline. Data: exp/tdmpc_glass/tamp_opencab/{res_eval_s1..7.json, RESULTS.json}. Prior: Part 40 (structured cap), Part 42 (curriculum beat), homepage Part 5 (the TAMP plan).