Keyboard shortcuts

Press ← or → to navigate between chapters

Press S or / to search in the book

Press ? to show this help

Press Esc to hide this help

Decoupled action expert: the denoiser was never the hard part

Read 2026-08-09 (lit slice lit-radar-0815, priority 3: the fjoint seam question). Paper: 2511.12101 — “Decoupled Action Expert: Confining Task Knowledge to the Conditioning Pathway” (Zhou, Lin, Fu, Li, Zhou, Wu — Australian Institute for Machine Learning, University of Adelaide; v1 2025-11-15, v2 2026-03-14, preprint, cs.RO). Read against the attachment decision and the fjoint pre-reg draft: the capacity half of the seam question — how much has to live in the expert vs arrive through conditioning — finally has a controlled measurement, though on Diffusion Policy, not a VLA.

The paper in plain words. A robot policy that generates arm motions by denoising has two jobs bundled together: knowing what the task wants (from camera images and instructions) and knowing how to shape a physically plausible 16-step arm trajectory. This paper argues the second job is tiny — a policy emits only 16×10=160 numbers per prediction, nothing like an image — and shows you can split the jobs cleanly: pretrain a small generic “motion” network on kinematics data that contains no images and no tasks at all (just joint-angles-to-gripper-pose sequences), freeze it forever, and teach each new task only to the side-channel that modulates it. The frozen generic core plus a retrained side-channel nearly matches full training, and a 5M-parameter MLP core matches — on their benchmarks, beats — the standard 244M U-Net. The catch is how the side-channel connects: modulation-style conditioning (scale/shift, as in AdaLN/FiLM) survives the freeze; cross-attention conditioning collapses to near zero, because task knowledge soaks into the frozen weights themselves.

What it contributes

  • A three-way factorization with the seam made explicit. Policy = observation encoders (ResNet-18/50) → conditioning network (produces per-layer FiLM γ/β) → action backbone (the denoiser). Decoupled recipe: Stage 1 trains {conditioning, backbone} on observation-free forward-kinematics data (joint positions → end-effector pose sequences, extractable from any trajectory dataset “at negligible cost”); Stage 2 freezes the backbone and trains fresh encoders + conditioning for the downstream task.
  • DP-MLP: the 5M backbone. Replaces Diffusion Policy’s 244M CNN U-Net with L residual FiLM-MLP blocks (h ← LN(h + W₂(γ⊙GELU(W₁h) + β))), a ~51× parameter cut. Two separable claims ride on it: (i) capacity — DP-MLP beats DP-C even under normal end-to-end training; (ii) decoupling — the frozen-backbone version keeps nearly all of it.
  • The conditioning-mechanism law. On one matched 8-layer transformer, freezing the backbone costs: cross-attention −41.5 pts, prefix tuning −36.8, additive −11.7, FiLM −7.7, adaRMSNorm −3.8, AdaLN +0.3, AdaLN-Zero +0.5. Token/attention routes embed condition-specific projections inside the backbone weights and collapse when frozen; modulation routes keep them outside and don’t. (They note π0.5 ships adaRMSNorm, GR00T-N1 AdaLN — the production experts are already on the decouplable side.)
  • The pretraining signal barely matters; pretraining does. A random frozen backbone scores 0.0. Unconditional, self-conditioned and joint-position-conditioned Stage-1 variants land within 1.6 pts of each other (62.2–63.8 MimicGen avg): the backbone learns generic trajectory structure, not task content.

The experiments it ran

MimicGen (8 tasks, 1000 demos each) and LIBERO (4 suites, 50 demos/task, DistilBERT language via the conditioning pathway), 3 seeds, max success rate. Headlines: LIBERO avg — DP-MLP normal 84.7 vs DP-C normal 79.3 (+10.9 on Long), and DP-MLP decoupled 84.2 (−0.5 from its own normal); MimicGen — DP-MLP normal 65.9 vs DP-C 63.6, decoupled 61.2 (−4.7, driven by Coffee/Stack3/Square). DP-T (cross-attention Diffusion Policy) under decoupling: 76.4 → 5.9 LIBERO — the mechanism law in benchmark form. Cross-embodiment Stage 1: pretraining the backbone on 76k external DROID Franka trajectories beats in-distribution FK pretraining (63.8/78.3 vs 62.2/76.8) and essentially closes the gap to normal training. Fine print: simulation only; Diffusion Policy only (“directly validating the decoupled recipe on full VLA systems remains important future work”); flow matching never tested; no inference-speedup or minimum-size sweep; decoupled DP-C still trails normal by 1.4–2.5 pts before the DROID rescue.

What transfers to us

The seam question (“how much capacity/task knowledge must live in the expert vs arrive through the conditioning pathway”) gets its sharpest capacity-axis datum, and it points one way: the denoising function class is tiny. Concretely for #4:

  1. The F arm is not capacity-starved, so the fjoint unfreeze shouldn’t be justified on capacity grounds. Our h1024×12 expert sits two orders of magnitude above the 5M floor this paper measures for the pure denoising job; whatever the frozen-trunk configuration lacks, it is not room in the expert. Sharpened prior for the fjoint rung: if J beats F2, read it as the trunk’s representations adapting (task-relevant features the frozen taps don’t surface), not as gradient relief for an overworked expert — and if J≈F2, this paper is the null’s mechanism (task knowledge was already arriving fine through a trainable conditioning path).
  2. Expert sizing (inherited h512/h1536 arm) gets a direction. The capacity result says smaller-with-good-conditioning is the live end of that dial, not larger — consistent with the ~11:1 trunk:expert ratios on the Hy-Embodied page, and cheap to check because sizing rungs were already priced as screens.
  3. Our expert checkpoint is task-entangled capital, not a generic head. Our expert conditions by cross-attention to residual taps — exactly the mechanism whose weights absorb condition-specific structure in their ablation. The F@10k expert is trunk-specific and task-specific; it warm-starts the fjoint rung (same trunk, byte-checked by materialize_fjoint_init.py) but should not be expected to survive a trunk swap (#17) the way a modulation-conditioned expert might. If a reusable-expert ambition ever appears, the seam mechanism — not the size — is what to change first.
  4. Observation-free Stage 1 is the cheapest expert-init yet filed. APT pretrained the expert on vision-action pairs with language masked; this paper gets a working prior from kinematics alone, no images, transferring across embodiments (DROID→MimicGen/LIBERO). Radar for any future fresh-expert attach: a near-free FK pretraining pass is now a published alternative to random init — the same random-init damage regime APT diagnosed, attacked from below.

What doesn’t transfer

  • The freeze is on the opposite side of the seam. They freeze the action backbone and retrain the conditioning; our F arm freezes the conditioning source (the trunk) and trains the expert. The result constrains the capacity split — it is not evidence for or against the frozen-trunk-vs-joint contrast the decision memo settled, and it does not re-rank F vs K.
  • Diffusion Policy, not a VLA; DDPM-style noise prediction, not flow matching; ResNet/DistilBERT conditioning, not a 4B VLM trunk. The authors chose DP precisely to strip the VLM out as a confound; the price is that nothing here measures what happens when the conditioning pathway is a frozen language model’s residual streams.
  • Sim-only, and the headline is partly a capacity story. DP-MLP normal already beats DP-C — some of “5M matches 244M” is “244M was oversized for these benchmarks,” a MimicGen/LIBERO fact, not a manipulation law.
  • The cross-attention collapse is a frozen-backbone fact. Our expert’s cross-attention taps are fine while the expert trains (both our arms train it); the −41.5 only bites if we ever freeze the expert and expect conditioning-side retraining to steer it.

Which idea/arm it fed

Idea #4 (seam-screen, decided — fjoint rung open): a capacity prior for the rung’s interpretation (J-wins ⇒ representation adaptation, not expert relief; J≈F2 ⇒ conditioning was sufficient), a direction for the inherited expert-width dial (down, not up), and the task-entanglement caveat on treating the F@10k expert as reusable capital across trunks. No gate, bar, or design change to the pre-registered rung. Cross-refs: APT (expert init from above), ActionX (the rung’s shape), seam-debate / π0.5-KI (the gradient side of the same seam), VLAFlow (recipe bake-off the capacity axis was missing from).