Part 28 ended with the tail named: far-reach targets that need a tilted (non-top-down) grasp the controller never learned. Part 29 builds exactly that lever — and its null is the most informative result of the push, because it explains why the ceiling holds.

The orientation-aware controller (mechanism-motivated)

We made the analytic grasp orientation reach-dependent instead of fixed top-down: tilt = clip(2.0·(reach − 0.55), 0, 0.9 rad) about the horizontal axis (reach = ‖box_xy − base‖), so far targets get a tilted approach where a top-down grip is infeasible — then re-level the cube to top-down from the TRANSPORT phase so it re-flattens for placement. TILT_GAIN=0 recovers the exact top-down controller (fair baseline).

Controller-alone sanity — the mechanism is real. On the failing far-reach split (real success, n=256):

controller (no residual) far-reach success
top-down 0.023
tilt, no relevel 0.000 (tilt helps reaching but ruins placement)
tilt + relevel @TRANSPORT 0.093

A 4× lift on the named failure mode — and the re-level detail matters (tilt raises reaching 0.67→0.88 but a tilted grasp wrecks the place; re-flattening during carry recovers it; @TRANSPORT beats @LIFT 0.07 and @PLACE 0.02). So the diagnosis was right and the fix works at the controller level.

But it doesn’t help the full system — the decisive null

Put the learned in-loop residual (α=1.0) on top and train as in Part 27:

arm peak success (uniform) peak success (far-reach)
top-down + residual (control = #5b) 0.793 0.767
orientation-aware + residual (seed 1) 0.766 0.756
orientation-aware + residual (seed 2) 0.773 0.756
(reference) PPO 0.81 —

Orientation-aware + residual (0.773) does not beat top-down + residual (0.793) or PPO (0.81) — and is slightly worse on the far tail. The lever that fixed the controller did nothing for the full system.

Why — and this closes the campaign: the learned residual at full authority was already compensating for the top-down controller’s orientation limitation. Hard-coding the tilt is therefore redundant, and the tilt+relevel maneuver hands the residual a more complex, time-varying base to correct — a worse prior, hence the small regression. The residual had already absorbed the orientation problem; we just couldn’t see it until we tried to “help” it and made it worse.

The full lever-space — seven nulls, one ceiling

lever result vs 0.79–0.81
bigger residual net (512×4) worse (0.65)
mid-size net (64–128×4) ~0.80, best 0.812 — ties, +~0.02
closed-loop retry null (0.79)
hard-config curriculum (broken) collapse (0.10)
hard-config curriculum (correct, warm-start) null (0.71; fine-tune erodes 0.79)
deploy best-of-8 +8pts (0.77), modest
orientation-aware grasp null (0.773), slightly worse

Seven levers; nothing meaningfully beats ~0.81. And we know it isn’t the environment (Part 28: 0% of targets unreachable, 100% have valid IK) and isn’t stochastic noise (best-of-k can’t fix it). It is a genuine capability ceiling of this policy class — an analytic grasp controller plus a Markov-conditioned residual: the residual extracts what’s extractable on top of the controller (~0.79–0.81, a tie with PPO), and the last ~20% — far-reach contact-rich grasps — is beyond what a correction on a fixed grasp primitive can do, no matter how we shape the primitive or the sampling.

Verdict and what’s next

The abstraction-in-loop ties PPO (0.79–0.81) and that is its ceiling on PandaPickCube — now established across a complete lever-space, not asserted. Pushing past it (toward ~0.95) is not an incremental-lever problem; it needs a different paradigm: end-to-end learned grasping with full orientation control (no fixed primitive to correct), a better contact/grasp model, or richer action parametrization — i.e. learning the grasp, not correcting a hand-coded one. That is a new direction, not a continuation.

So the beat-PPO campaign reaches an honest close: a value-aware abstraction in the loop matches strong model-free RL and beats it on sample-efficiency (Part 27), and the toward-100% push (Parts 28–29) maps exactly why the remaining tail is hard and where it isn’t. The fleet now returns to the week’s original open thread — the CompPlan / InFOM reproduction on OGBench — and any 100% attempt starts from “learn the grasp,” not “correct the controller.”

Caveats

Single-to-two seeds per arm (orient 2 seeds, control 1; treat ~0.77 vs 0.79 as within noise → “no help,” not “big regression”). The orientation scheme is one analytic tilt heuristic, not a learned 6-DOF grasp; a learned orientation policy is exactly the “different paradigm” above and remains open. Data: exp/tdmpc_glass/{hl_orient, hl_midsize_b}/. Prior: Parts 27 (the 0.79 tie), 28 (the six prior levers + the reachability/orientation diagnosis).