Part 27 left the abstraction-in-loop at 0.79, a near-tie with PPO (0.81), with ~20% of configs failing. This week we threw six experiments at that tail to chase ~0.95/100%. None breaks it — but the diagnostics tell us exactly what the tail is, which is worth more than another null.

The six levers (real success, box_target≥0.9, n=256)

# lever result verdict
#6 bigger residual net (512×4) 0.65 hurts — over-capacity for a residual that needs small corrections
mid mid-size sweep (64–128×4) mean ~0.80, best 0.812 mild sweet-spot — ties PPO’s 0.81
#7 closed-loop retry (re-grasp in-episode) 0.79 null — fires up to 4.9 attempts/ep, no gain
#8 hard-config curriculum (as first built) 0.10 broken (from-scratch + 50% hard from step 0)
#8b hard-config curriculum done right 0.71 null — and fine-tuning erodes the 0.79 optimum
#11 deploy-time best-of-8 (no retraining) 0.77 uniform modest +8pts, recovers ~21% of hard configs
— (reference) PPO / our #5b α=1.0 0.81 / 0.79 the bar

Capacity has a sweet-spot. 32×4 → 0.79, 64–128×4 → ~0.80–0.81 (best seed 0.812), 512×4 → 0.65. So the residual wants a little more capacity than the tiny default but collapses if it’s too big — and at the sweet-spot it cleanly ties PPO’s peak. That’s the only lever that moved the number up, and only by ~0.02.

Everything else is a null, and #8b adds a sharp sub-finding: the 0.79 policy is a fragile optimum — any continued fine-tuning (curriculum or uniform control, identical full-reset regime) makes both arms peak at the carried-in checkpoint and then decay monotonically. So you can’t fine-tune your way past 0.79; prioritized hard-config sampling does nothing on top of that.

The diagnostics — what the 20% tail actually is

Two cheap experiments did more than all the training levers combined.

(#10) Is ~0.95/100% even reachable? — Yes, by construction. Building the Panda’s FK workspace (300k joint samples) + analytical IK:

  • 0% of targets are geometrically unreachable (max reach 0.947 m; the hardest mined failure sits at 0.919 m). 100% of uniform targets and 100% of the 1160 mined hard-failures have a valid, in-limit IK solution. So the environment is not the ceiling — ~0.95+ is attainable in principle.
  • The difficulty is orientation, not position. Only 13.6% of the hard-failure configs admit a strict top-down grasp, versus 66% of uniform configs. Failures cluster at far horizontal reach (48% beyond 0.85 m; mean target x = 0.82 vs 0.70 box center), where the arm must tilt and reorient the gripper to grasp at all.

(#11) Are the failures stochastic or deterministic? — Mostly deterministic. Best-of-K (the 0.79 policy, K independent stochastic rollouts, solved if any succeeds):

K uniform hard configs
1 0.69 0.12
8 0.77 0.21

Best-of-8 recovers ~21% of the hard tail (those are stochastically solvable) but ~79% stay unsolved across all 8 rollouts — deterministic failures. More sampling won’t close it.

The verdict — and the named next lever

0.79–0.81 is a robust ceiling for this policy class — a top-down-grasp analytic controller plus a Markov-conditioned residual. It ties PPO and no capacity/curriculum/retry/sampling lever pushes meaningfully past it. But the ceiling is not the environment (everything is reachable) and not stochastic noise (best-of-k can’t fix it). It is one specific, named policy limitation:

The controller only does top-down grasps; the ~20% tail is far-reach targets that require a tilted, reoriented grasp the policy never learned.

That is the next experiment, and the first one motivated by a mechanism rather than a hope: a grasp-orientation-aware controller/residual — let the grasp phase choose an approach orientation as a function of the target’s reach, so far-reach configs get a tilted grasp instead of an impossible top-down one. Combined with deploy-time best-of-k (+8pts), that is the credible path from 0.81 toward ~0.9+.

Caveats (kept honest)

Single-to-two seeds per lever (the 0.812 mid-size peaks are two independent seeds; treat as ~0.80 ± noise, a tie-not-a-beat vs PPO 0.81). Reachability uses analytical IK with an orientation set, not a full grasp-physics test (a target with a valid IK can still be hard to grasp stably — consistent with “hard-but-solvable”). best-of-k uses the policy’s own stochasticity. Data: exp/tdmpc_glass/{hl_beatppo_big, hl_midsize, hl_retry, hl_curriculum_b, hl_tail}/. Prior: Parts 24–27.