Does a Learned Corrector Beat a Simple Retreat? Evidence from a Frozen VLA

Chenchao Sheng; Zhuang Jiang; Liuhaichen Yang; Ningwei Bai; Zezhi Tang

cs.ROOct 2, 2026

Does a Learned Corrector Beat a Simple Retreat? Evidence from a Frozen VLA

Authors: Chenchao Sheng, Zhuang Jiang, Liuhaichen Yang, Ningwei Bai, Zezhi Tang

Organizations: University College London (UCL), London, United Kingdom

Abstract

Before deploying runtime recovery for a frozen vision-language-action (VLA) policy, one must establish that an intervention improves success beyond ordinary run-to-run variation and that its complexity adds value over a simple action. We evaluate these questions on frozen $π_{0.5}$ across four RoboTwin tasks. For each test seed, we pair rollouts with and without correction and include a same-seed base-policy re-run as a placebo. Seed-cluster intervals and prespecified comparison rules assess net gains against stochastic outcome changes. Across 3,888 paired episodes, the full pipeline raises success on beat_allowbreak block_allowbreak hammer by $+13.5$ ,pp (95% interval $[+9.4,+17.7]$ ), with no detectable gain on the other three tasks at the deployed weight. Among failed base episodes on the responsive task, $43.2\%$ succeed on a plain re-run, compared with $63.5\%$ after correction; many nominal rescues therefore reflect the base policy's own variability. A fixed-time trigger and scripted return to an earlier joint configuration produce a net gain with no detected difference from the learned pipeline across two rounds, although our prespecified equivalence criterion is not met consistently. Pausing and a constant-action control do not yield comparable gains. On this benchmark, the decision to intervene depends strongly on the task, and a paired placebo plus a simple retreat baseline are needed to establish what learned correction contributes.

Figures & tables

Figure 1: The pipeline under test and the protocol that measures it. Left: the frozen policy, the stall monitor and the distilled corrector run at deployment; the privileged teacher and the counterfactual admission step run only on training seeds, where 87 of 175 candidate rescues are kept. Right, bbh: 63.5% of failed base episodes succeed under the intervention against a placebo of 43.2%, leaving $+20.4$ pp; below that, net effects for three conditions with 95% intervals, then the drift-cancelling difference between retreat and corrector, whose 90% interval (not drawn) leaves the $\pm 5$ pp band (Section 3.3 ). Bottom: one bbh test seed, the same alarm at step 149, under the scripted retreat and under the corrector; four frames each, in order and not to scale.

Task	$n$	$b$ (%)	Net (pp) [95% CI]	$p$	$\rho/\rho_{0}$ (%)	$\beta/\beta_{0}$ (%)	Placebo-corr. rescue / breakage
bbh (beat_block_hammer)	972	54.0	13.5 [9.4, 17.7]	$<10^{-4}$	63.5 / 43.2	29.1 / 36.8	20.4 [14.1, 26.2] / $-$ 7.6 [ $-$ 12.9, $-$ 2.5]
qr (rotate_qrcode)	972	60.8	0.6 [ $-$ 2.9, 4.2]	0.77	40.7 / 43.3	25.2 / 27.9	$-$ 2.6 [ $-$ 8.8, 3.4] / $-$ 2.7 [ $-$ 7.2, 1.7]
a2b (place_a2b_left)	972	38.3	2.3 [ $-$ 1.0, 5.6]	0.18	22.5 / 20.8	30.4 / 33.6	1.7 [ $-$ 2.1, 5.6] / $-$ 3.2 [ $-$ 8.9, 2.2]
a2br (place_a2b_right)	972	34.0	$-$ 0.8 [ $-$ 4.3, 2.5]	0.68	21.5 / 21.5	44.2 / 41.8	0.0 [ $-$ 4.2, 3.9] / 2.4 [ $-$ 3.7, 8.3]
pooled	3,888	46.8	3.9 [2.1, 5.7]	$2.3\times 10^{-5}$	34.4 / 30.0	30.9 / 34.2	—
pooled without bbh	2,916	44.3	0.7 [ $-$ 1.4, 2.7]	0.51	26.4 / 26.4	31.6 / 33.1	—

Table 1: Main table. Four RoboTwin tasks, 972 paired episodes each (324 seeds

\times

3 repeats).

b

: base success. Net: intervention minus base success (pp), with a 95% seed-cluster bootstrap interval and McNemar

p

\rho/\rho_{0}

and

\beta/\beta_{0}

: rescue and breakage rates with their same-seed placebo. Placebo-corrected rescue

=\rho-\rho_{0}

, breakage

=\beta-\beta_{0}

Figure 2: The deployed corrector, and the ladder on bbh. Filled markers exclude zero throughout. (a) Net change in success rate per task and pooled, with each task’s base rate

b

. (b) Placebo-corrected rescue and breakage. The upper line of each pair, marked

\rho

, is

\rho-\rho_{0}

; the lower, marked

\beta

, is

\beta-\beta_{0}

. The two rates have different denominators and do not add. The privileged-planner row is blank here because that condition has no same-seed placebo, and it carries no interval in (c) because its per-episode pairs were not saved. (c) The ladder: six interventions sharing one trigger threshold, with 95% seed-cluster bootstrap intervals where available. The three retreat and pipeline rows run on 972 pairs, the rest on 324.

Figure 3: Replication round: four conditions on bbh, 972 pairs each, on one GPU in one time window. (a) Net change in success rate, with each arm’s base rate

b

; filled markers exclude zero at 95%. (b) Pairwise differences, five within the round (90% intervals, upper group) and three between rounds (95%, lower group), against a shaded

\pm 5

pp margin; pale red marks “no difference detected”, and the one dark marker is the pair of scripted conditions, the only comparison here that is equivalent. Every one of the eight contains zero, and the between-round net effects differ by at most 1.5 pp.

Ablation	bbh	qr	a2b	pooled	Verdict
Admission: all 175 vs admitted 87	$-$ 1.1 [ $-$ 6.0, 3.9]	0.5 [ $-$ 3.9, 4.8]	$-$ 2.5 [ $-$ 6.6, 1.5]	$-$ 1.0 [ $-$ 3.6, 1.5]	equivalent
Admission: random 87 (draw a) vs admitted 87	$-$ 4.2 [ $-$ 9.3, 0.9]	$-$ 0.6 [ $-$ 4.9, 3.7]	$-$ 2.8 [ $-$ 6.8, 1.2]	$-$ 2.5 [ $-$ 5.1, 0.1]	equivalent
Admission: random 87 (draw b) vs admitted 87	0.3 [ $-$ 4.3, 4.9]	$-$ 0.6 [ $-$ 5.1, 3.9]	$-$ 5.1 [ $-$ 9.0, $-$ 1.2]	$-$ 1.8 [ $-$ 4.3, 0.7]	equivalent
Trigger: timer vs monitor	$-$ 2.0 [ $-$ 6.8, 2.9]	0.3 [ $-$ 3.9, 4.5]	$-$ 4.8 [ $-$ 8.5, $-$ 1.2]	$-$ 2.2 [ $-$ 4.6, 0.3]	equivalent
Teacher: rung 1 only (60 rescues)	$-$ 4.3 [ $-$ 9.6, 0.9]	$-$ 2.8 [ $-$ 7.1, 1.5]	0.0 [ $-$ 4.0, 4.0]	$-$ 2.4 [ $-$ 5.0, 0.3]	equivalent
Hand-back depth: 20 steps vs 40	$-$ 5.4 [ $-$ 10.6, $-$ 0.2]	4.0 [ $-$ 0.2, 8.2]	$-$ 3.1 [ $-$ 7.1, 0.8]	$-$ 1.5 [ $-$ 4.1, 1.1]	equivalent

Table 2: Component ablations. 648 paired episodes per task; control is a three-task corrector trained on the 87 admitted rescues (40-step hand-back). Entries are variant minus control in intervention success, with 95% intervals. Equivalence was preregistered for the pooled effect only; per-task columns carry no equivalence verdict. Verdict column: equivalent = 90% pooled interval inside

\pm 5

pp; worse = 95% pooled interval excludes zero and is negative.

Figure 4: Rescue–breakage plane. Circles: the deployed corrector, 972 pairs per task. Squares: the scripted retreat, 972 pairs on bbh (monitor + retreat) and 324 on the others (retreat). Open grey diamonds: the same-seed base re-run, the placebo; arrows run from each placebo to its corrector. The single dark marker is the privileged planner online, bbh only, 324 pairs and no placebo. Dotted: the break-even lines

\rho/\beta=b/(1-b)

, one per task, each labelled with its task and that task’s base rate

b

from the corrector arm. Only bbh sits well above its line under both interventions.

Figure 5: Component ablations: each variant minus the three-task control, 648 pairs per task. Left: per task, shaded by how much worse the variant is, so an unshaded cell is an improvement; boxed where the 95% interval excludes zero, and the intervals themselves are in Table 2 . Right: pooled, filled where the interval excludes zero; bars are 95% intervals. The shaded band is

\pm 5

pp: all eight pooled estimates fall inside it and two are graded worse. Equivalence verdicts use 90% intervals, which are not drawn (Section 3.3 ).

Figure 6: Sensitivity of the bbh effect to input noise and output delay, 972 pairs per condition. Top: net gain at four noise levels on the joint angles the corrector reads and at a four-step delay on its output; the leftmost point, at zero noise, is the clean reference. The noise ticks are not to scale. Bottom: the endpoint difference between each condition and the clean one, with 95% intervals, not the gaps between the points above, which are differences of net effects; the grey marker at zero noise is the clean condition against itself. All five contain zero, and the largest noise level has a bound at exactly zero. Two of the five exclude zero on their 90% intervals, which are not drawn. Read as differences of net effects, the same five span

+0.3

-5.3

pp (Section 3.3 ).

Robot Failure Recovery

Robot failure recovery studies how robots detect, diagnose, and respond to errors, damage, and unexpected physical interactions.

See all 73 papers

If at First You Don't Succeed, Backtrack: Lightweight Recovery for Frozen VLAs
Oct 23, 2026Thomas Landeg, Thierry TambeVision-Language-Action Models Robotic Manipulation
Stanford University · ETH Zurich
Retrieve in Time, Correct in Frequency
Aug 5, 2026Yuze Fan, Yue Cao, Pengjie Gao +7Robotic Manipulation Vision-Language-Action Models
Research Institute of Tsinghua University in Shenzhen · Tongji University · Everwise-Tech Co., Ltd. +2
Retrieval-triggered Action Projection Adjustment for Frozen VLA Policies: An Empirical Study and Failure Analysis
Oct 17, 2026Jinming Wang, Ruiheng Zhang, Yong Wang +2Vision-Language-Action Models VLA Model Adaptation
Department of Computer Science and Technology, Hainan Bielefeld University of Applied Sciences, Danzhou, China

Oct 23, 2026J/K move · Enter open · S save

If at First You Don't Succeed, Backtrack: Lightweight Recovery for Frozen VLAs

Thomas Landeg, Thierry Tambe

Stanford University · ETH Zurich

Vision-Language-Action Models Robotic Manipulation

Open paper AskPDF arXiv

Vision-Language-Action (VLA) models perform well on robotic manipulation, yet remain brittle once their own actions carry them into states their training data barely covers. Recovering from such failures usually requires recovery demonstrations, policy retraining, or a separate model at inference time. We introduce Backtrack, a lightweight failure detection and recovery mechanism for frozen VLAs that needs none of these. A small failure prediction head, trained only on outcome-labeled rollouts of the frozen policy, reads the VLA's existing internal representations to estimate whether the current episode will fail. When the predicted failure probability exceeds a threshold, a controller pauses the policy, retraces the arm's recently achieved motion, and hands control back to the unchanged policy so it can retry from an earlier configuration. The head adds only 0.48-0.88% to the policy's parameter count. We evaluate Backtrack on LIBERO with three VLAs spanning two action-generation families: SmolVLA and

π

0.5, which use flow-matching action experts, and the autoregressive

π

0-FAST. On LIBERO-Long, the suite of long-horizon multi-step tasks and the one on which all three policies fail most often, Backtrack reduces each policy's failure rate by 19-48%, raising success by up to 7.4 percentage points; on the three shorter suites, where control success is already 85-99%, its effect is smaller. On MetaWorld, all twelve controller configurations tested with SmolVLA improve on the unassisted policy.

Abstract

Robot Failure Recovery

Explore similar work

If at First You Don't Succeed, Backtrack: Lightweight Recovery for Frozen VLAs

Retrieve in Time, Correct in Frequency

Retrieval-triggered Action Projection Adjustment for Frozen VLA Policies: An Empirical Study and Failure Analysis