The state shortcut — proprioception as a crutch, and what we measured when we kicked it away

Papers: Adapt Your Body (2506.23944, withdrawn), the state-free policy (2509.18644), ReViP (2601.16667), GAP (2602.12032, ICLR 2026), ThinkProprio (2602.06575), and Cloak (2606.22836). Banked across the 2026-08-06 lit slices; re-read at full-text depth for this page. Fed: #9 (state dropout — pre-registered, run, and falsified at p=0.8 on our stack) and #11 (the grounding gap / reliance probe, where the mechanism was confirmed). This is the rare theme where we can score the literature against our own completed experiments — and where the retroactive deep read caught our banked citation being wrong in a way that matters.

The theme

Imitation-trained manipulation policies get two views of the world: cameras, and the robot’s own joint state. State is low-dimensional, noiseless, and — because expert demos are smooth — an excellent predictor of the next action from the current one. That makes it a classic shortcut feature: fast to learn, great on the training distribution, and a liability the moment reality diverges from the demo manifold. The papers here all orbit one question: how much of a policy’s competence is visual understanding, and how much is dead-reckoning off proprioception? Their answers range from “remove state entirely” to “rebalance it adaptively” — and they disagree in an instructive way.

What we ran — both sides of the story

This literature fed two completed experiments here, which came back with a spliced verdict.

The mechanism is real (reliance probe, #11). We froze a masked evaluation (--mask-state substitutes the dataset state mean) and asked whether B — the arm with the better intact first-action error — was buying it with heavier state reliance. It was: D = Δ_first(B) − Δ_first(A-s0) = +0.702, CI95 [0.498, 0.916], 14× the pre-registered threshold, the chunk secondary agreeing (+0.389). Exactly the causal-confusion story: the better leaderboard number was partly a deeper draw on the proprioceptive crutch.

The naive fix did not pay (state dropout, #9). The supposedly literature-endorsed lever — zero-masking state with p=0.8 — went into a pre-registered 40k arm (--state-dropout 0.8). Outcome: paired Δchunk +2.64 MAE [2.55, 2.74], far outside the ±0.15 band — a clear COSTS verdict. But the mechanism worked exactly as advertised: the dropout-trained model barely notices state masking (first_mae 8.56 intact → 8.11 masked) while the baseline collapses under the same mask (3.94 → 24.08). Our one-sentence summary: dropout kills the shortcut without teaching the replacement. A p=0.3 screen stays queued; nothing was adopted.

The deep read then re-framed both results — see the first paper.

1. Adapt Your Body — our citation was wrong, and the paper is withdrawn (2506.23944, v1)

The correction first: p=0.8 zero-masking is not this paper’s method. It is their baseline — “Random Dropout: randomly mask proprioception input to zeros with probability 80%” — which their actual method beats. The method (NADA) is two-pass: train a policy, roll it out, measure a time-conditioned Wasserstein distance between expert and rollout state distributions, pick the noise scale σ* that minimizes it, and retrain with Gaussian-noised proprioception. On Robomimic/MimicGen (9 sim tasks, 3 seeds) NADA posts e.g. Square 56.7 vs dropout’s 44.8, StackThree 61.4 vs 52.7 — best on 6/9 tasks — with the optimal σ visibly task-dependent (0.6 to 1.2 on their grid). Their per-dimension analysis fingers velocity terms as the most harmful proprio components. And then the kicker: v2 of this paper is a withdrawal notice — the only content version is v1, reason unstated.

So our #9 arm C was, strictly, a test of this paper’s straw man at its quoted strength — and the straw man cost us +2.64 MAE, the same direction their own table points (masking is dominated by calibrated noise). The honest bank update: the input-corruption family is still live, but its literature-backed form is noise scaled to the deployment gap, not heavy zero-masking, and its flagship citation is withdrawn evidence. The queued p=0.3 screen survives on our own branch rule, not on this paper’s authority.

2. The state-free policy — amputation works, but only with two accomplices (2509.18644, v2)

The maximalist answer: drop state entirely. Confirmed as banked — but the full text shows the vision-only headline rests on two specific enablers, not on deletion alone: (a) relative end-effector actions (displacements from the current pose), and (b) “full task observation” — dual wide-angle wrist cameras (~120°×120°). With both, real-robot spatial generalization goes from ~0 to 0.98 (height) / 0.58 (horizontal) on Pick Pen, and the pattern is architecture-agnostic (π0, ACT, Diffusion Policy all flip from ~0 to strong). Without the action-space change, everything scores exactly zero — relative-joint, absolute-joint, absolute-EE all dead on generalization. Their rescue attempts for state-based policies (noise aug 0.633, diverse data 0.117, LoRA 0) all lose to state-free 0.983; an overhead camera actively hurts extreme generalization. In-domain, state-based and state-free tie (LIBERO 0.938 vs 0.945).

For us: the cleanest external statement that state removal is a system choice, not a flag. Our arm C removed the crutch (their diagnosis held) but had neither accomplice — our action space is absolute joint-style chunks and our cameras are what the corpus gives us — and their own ablation table predicts exactly what we measured: zeros without the enablers.

3. ReViP — “false completion” and the case for modulating instead (2601.16667, v3)

The diagnostic paper. “False completion”: a policy acts as if the goal is achieved because its internal progress estimate (carried by proprioception and habit) says so, ignoring visual evidence — their motivating real-robot study shows a state-enabled π0 with 70% success still false-completes on 46/50 perturbed trials; naive state masking cuts false completion to 17/50 but drops success to 40% — the same modulate-don’t-amputate lesson as our arm C, from an independent lab. Their fix keeps state and rebalances: an external VLM (Qwen2.5-VL-3B) extracts progress-aware visual cues, which FiLM-modulate the vision-language prefix before action generation. On their 8-task perturbation benchmark (objects dropped mid-episode, distractors swapped, scenes relaid): π0 36% → ReViP 59% (+23; the banked “+26” is ReViP* with a 72B VLM — a precision fix to our note). Real-robot 62% → 88%. Cost: 44.6 → 62.4 ms per step, and a second model in the loop.

For us: the mechanism story matches our probe result almost line-for-line (better intact numbers riding on state reliance), and it names the failure mode our offline panel can’t see — false completion is a rollout phenomenon. Banked as the heavier modulation-side arm behind dropout in #9’s original design; the deep read keeps it there but upgrades its evidence class: real-robot confirmed, ablated across VLM backbones.

4. GAP — the training-dynamics cause, with the receipts (2602.12032, v1, ICLR 2026)

The why. GAP runs temporally controlled interventions — splice the vision-proprio policy’s actions into a vision-only rollout for 10-step windows — and finds degradation concentrates in motion-transition phases (locate/reorient moments: −7 to −14%), nowhere else. Cause: proprio’s concise signal gives faster early loss reduction, winning the modality competition and suppressing visual-branch gradients precisely where vision matters. The fix is optimizer-side: segment the demo into motion-consistent phases (change-point detection on proprio deltas, LSTM smoothing into a transition probability ρ), then shrink the proprio branch’s update by λ(1−ρ) during the first half of training only. Verified numbers: assembly 74.6 (concat) → 94.2, threading 33.2 → 53.0; real dual-arm lift-lid-and-pour 5/20 → 15/20; on a fine-tuned Octo it averages +17% over the vision-proprio baseline — though “works on VLAs” rests on that one VLA. A fixed-probability masking baseline (RDT-style) loses to GAP everywhere it’s reported.

For us: still the named follow-on if input-side corruption plateaus — now with the sharper framing that our arm C tested the family GAP’s own baseline column shows is dominated. GAP also made a testable side prediction we banked: the grounding gap should concentrate in motion-transition frames, readable for free from the probe npz by conditioning Δ_first on progress-within-episode.

5. ThinkProprio — state as a lens, not a crutch (2602.06575, v2)

Our thin note (“proprio as text tokens at the prompt input”) was directionally right but missed the actual contribution. Yes, each proprio scalar is binned to 256 levels and mapped into the pretrained vocabulary embedding table (no learned projection, Florence-2 backbone). But tokenized proprio alone buys ~nothing (CALVIN 4.45 vs baseline 4.44). The headline mechanism is state-grounded visual token selection: language and proprio tokens form two guidance branches that select ~12% of visual patches before the VLM runs. Both branches matter — language-only selection scores 3.40, proprio-only 3.12, together 4.52. The load-bearing ablation for the placement question we care about (#11): MLP-encoded proprio injected at the VLM hurts (4.15 vs 4.44 with none) while vocab-token proprio doesn’t — evidence that how state enters is as important as whether. CALVIN ABC→D 4.52 (FLOWER 4.44), LIBERO 97.7% avg, real UR3 88.9% vs FLOWER’s 80.7% at 22 ms vs 52 ms per step.

For us: the most constructive paper of the set — state used to direct visual attention rather than to predict actions. It inverts the shortcut: proprioception decides where to look, vision decides what to do. Banked into #11’s conditioning-placement discussion (our soft-state-token enters late; their table says late

• learned-projection is the worst quadrant they measured).

6. Cloak — masking on the vision side (2606.22836, v1)

The other masking axis: hide the end-effector from the model, visually, so the policy can’t bind to embodiment appearance — zero-shot cross-embodiment transfer follows. Mechanism (sharper than our note): not inpainting — the EE mask is rasterized geometrically from the robot model + state + camera parameters and applied as a ViT attention mask, with training-time augmentations (rolled-image fill, capsule attachment, disk removal) so the silhouette can’t be memorized; a generative model never touches the observations. Trained on DROID (Franka), zero-shot to a UMI gripper, a YAM arm, and a five-fingered Sharpa hand: task progression rate (their metric — stage completion, not binary success) 85.1/86.3/81.8 vs ~55–70 for π0.5-droid with tip-pose retargeting. Source-embodiment performance is uncompromised (88.0 vs 89.3, within error). Limits they admit: two-fingertip skills only, no rich contact.

For us: not actionable at the panel stage (we have no cross-embodiment axis yet), but it’s the standing answer to a north-star question — when the owner rig arrives with a different arm than the corpus embodiments, visual EE masking is the zero-shot lever with real evidence, and it composes with anything above (it’s vision-side; the state-side debate is orthogonal).

What transfers, what doesn’t, and what it fed

Transfers: the diagnosis, fully — measured here twice (probe D = +0.702; baseline collapse 3.94 → 24.08 under masking). And the cross-paper consensus the retroactive read surfaced: modulate, don’t amputate — ReViP’s masking study (success 70% → 40%), GAP’s dominated masking baseline, and Adapt Your Body’s own tables all point the same way, and our arm C is an independent fourth datapoint. Amputation works only in the state-free paper’s full system, with relative EE actions and wide-FOV wrist cameras doing half the work.

Doesn’t transfer: the p=0.8 recipe as literature-endorsed practice — that was a baseline in a withdrawn paper, mis-banked by our skim as the method. Success-rate margins on scripted perturbation suites don’t map to our offline chunk-MAE panel (false completion is invisible to it). ThinkProprio’s selector assumes a token-budget architecture we don’t run today.

Fed: #9 — arm C’s verdict re-framed (we falsified the family’s weakest member; calibrated noise à la NADA and GAP-style gradient scaling are the surviving levers, p=0.3 screen still queued on our own branch rule). #11 — ThinkProprio’s placement ablation joins the conditioning-placement evidence; GAP’s motion-transition prediction stays a free conditional read on the probe npz. North star — Cloak banked as the cross-embodiment lever for the rig. And one bank hygiene note: the ideas.md hook for 2506.23944 now carries the withdrawn flag and the baseline-not-method correction.