Learning from Runtime Feedback through Failure-Bank Self-Evolution for Vision-Language-Action Models

Mingyue Cui; Zheyuan Liu; Yihan Zhu; Zheyuan Zhang; Meng Jiang

cs.ROSep 30, 2026

Learning from Runtime Feedback through Failure-Bank Self-Evolution for Vision-Language-Action Models

Authors: Mingyue Cui, Zheyuan Liu, Yihan Zhu, Zheyuan Zhang, Meng Jiang

Organizations: University of Notre Dame

Abstract

Vision-language-action (VLA) models generalize broadly across robotic manipulation tasks, but complex environments require balancing task success with unintended contact. Runtime shields can correct individual actions, but they leave the underlying policy unchanged, so repeated disagreements may create a persistent policy-shield mismatch that blocks task progress. To address this challenge, we introduce FailBank, a four-stage self-evolving framework that converts runtime feedback into persistent policy improvement. During collection, a fixed CBF-based safety module serves as an observe-only teacher, producing counterfactual corrections while the policy remains in control. Outcome-aware admission then converts useful proposals into corrective targets and retains successful uncorrected actions as quiet anchors for guarded LoRA updates. We evaluate FailBank on the VLA-Arena benchmark across two difficulty levels and two VLA backbones. Compared with the base policies, FailBank improves the joint success-cost operating point. Across the two backbones, FailBank improves task success rate by 8.5 and 6.9 percentage points, while reducing policy-induced cumulative cost by 35.6% and 23.8%, respectively. Compared with runtime shielding, FailBank raises task success rate by 25.4 and 9.5 percentage points, while maintaining comparable policy-induced cumulative cost. These results show that runtime feedback can serve as persistent policy supervision rather than only as a temporary action constraint.

Figures & tables

Figure 1: Representative Level-2 static-obstacle results on VLA-Arena. Top: task success rate (SR,

\uparrow

better); bottom: policy-induced cumulative cost (

\mathrm{CC}_{\mathrm{policy}},\downarrow

better). AEGIS reduces cost while largely preserving success on Mango, fails to improve success on Apple, and sharply reduces success on Tomato. FailBank achieves the highest success rate across all three tasks while reducing policy-induced cost relative to Base.

Figure 2: A shield-collapse trajectory on the Level 2 onion task. AEGIS repeatedly intervenes near the bottles and eventually times out. FailBank uses observe-only CBF supervision and successfully places the onion in the bowl.

Figure 3: The four-stage FailBank loop. Stage 1 collects policy-controlled rollouts while logging CBF proposals without altering execution. Stage 2 discards invalid records and filters valid records into CBF-triggered corrections and quiet anchors, assigning their targets and weights. Stage 3 accumulates the valid records in the training bank. Stage 4 fits a fresh LoRA adapter and accepts the candidate policy only if it passes the held-out guard. The accepted policy is then used to collect rollouts in the next round.

Figure 4: Success–cost trade-off relative to the base policy. Circles and diamonds denote

\pi_{0.5}

and

\pi_{0}

, blue and orange denote FailBank and AEGIS, and shade indicates difficulty level. Upward movement means higher success, and leftward movement means lower policy-induced cost. Thus, the shaded upper-left quadrant is better on both. FailBank places more points in this joint-improvement region, while AEGIS more often moves left but downward.

		$\pi_{0.5}$									$\pi_{0}$
		Base			AEGIS			FailBank			Base			AEGIS			FailBank
Level	Task	SR $\uparrow$	CC $\downarrow$	$\mathrm{CC}_{\mathrm{policy}}\downarrow$	SR $\uparrow$	CC $\downarrow$	$\mathrm{CC}_{\mathrm{policy}}\downarrow$	SR $\uparrow$	CC $\downarrow$	$\mathrm{CC}_{\mathrm{policy}}\downarrow$	SR $\uparrow$	CC $\downarrow$	$\mathrm{CC}_{\mathrm{policy}}\downarrow$	SR $\uparrow$	CC $\downarrow$	$\mathrm{CC}_{\mathrm{policy}}\downarrow$	SR $\uparrow$	CC $\downarrow$	$\mathrm{CC}_{\mathrm{policy}}\downarrow$
1	Apple	90.0	9.12	0.12	82.0	8.20	0.00	90.3	9.21	0.21	54.0	5.40	0.00	82.0	8.20	0.00	59.0	5.90	0.00
	Lemon	83.0	44.07	35.87	78.0	7.81	0.01	81.0	25.98	17.88	66.0	6.60	0.00	80.0	8.00	0.00	84.0	8.40	0.00
	Mango	77.0	74.08	66.38	68.0	21.08	14.28	93.7	23.12	13.76	92.0	11.34	2.14	82.0	9.40	1.20	93.0	12.39	3.09
	Onion	72.0	33.88	27.18	44.0	9.12	4.71	90.0	31.63	23.50	28.0	18.34	16.74	30.0	3.00	0.00	24.0	12.02	10.32

Table 1: Static-obstacle results across two backbones. Collection happens in Level 1 Mango task while other tasks are unseen during rollout collection. SR is task success rate (%). CC is official cumulative cost.

\mathrm{CC}_{\mathrm{policy}}

is policy-induced cost. Best SR,

\mathrm{CC}_{\mathrm{policy}}

and BRS values are in bold .

Figure 5: Ablations of the collection interface and learning signals. (a) Paired episode outcomes under observe-only and shield-in-loop collection on Level 1 mango task. Rows show observe-only outcomes, while columns show outcomes when shield corrections are executed. (b) Failure records produced on the same task by the two collection modes, separated by policy failures and shield-induced failures. (c–d) Comparison of Base, training on CBF-triggered records (Correction-only), and training with additional successful uncorrected actions (+ quiet anchors), reporting SR and

\mathrm{CC}_{\mathrm{policy}}

on Level 1 onion task, respectively.

Figure 6: Self-evolution across accumulated rounds on Level 1 onion task. The failure bank grows from

3.7

k records at R1 to

22.7

k at R5. SR peaks at R2, whereas

\mathrm{CC}_{\mathrm{policy}}

reaches its minimum at R3. Later rounds fluctuate while remaining improved over base policy on both axes.

Appendix figures & tables31 assets

Supplementary material from the paper’s appendix.

Appendix

Safety / Static obstacles: pick the named object and place it in the bowl or plate
			Base			AEGIS			FailBank
Backbone	Level	Task	SR $\uparrow$	CC / $\mathrm{CC}_{\mathrm{policy}}$ $\downarrow$	BRS $\uparrow$	SR $\uparrow$	CC / $\mathrm{CC}_{\mathrm{policy}}$ $\downarrow$	BRS $\uparrow$	SR $\uparrow$	CC / $\mathrm{CC}_{\mathrm{policy}}$ $\downarrow$	BRS $\uparrow$
$\pi_{0.5}$	1	T0 Apple	90.0	9.12 / 0.12	0.368	82.0	8.20 / 0.00	0.407	90.3	9.21 / 0.21	0.261
		T1 Lemon	83.0	44.07 / 35.87	0.368	78.0	7.81 / 0.01	0.524	81.0	25.98 / 17.88	0.446
		T2 Mango	77.0	74.08 / 66.38	0.368	68.0	21.08 / 14.28	0.448	93.7	23.12 / 13.76	0.786
		T3 Onion	72.0	33.88 / 27.18	0.368	44.0	9.12 / 4.71	0.337	90.0	31.63 / 23.50	0.543

Appendix

Table 2: VLA-Arena static-obstacle results across two backbones. Cost cells report official

\mathrm{CC}

followed by diagnostic

\mathrm{CC}_{\mathrm{policy}}

. BRS summarizes base-relative failure and policy-induced cost. Dashes denote undefined BRS when the base policy-induced cost is zero. FailBank averages three training orders on

\pi_{0.5}

and the two of three that pass the training guard on

\pi_{0}

. L1-T2 is the collection task. L2-T1 is a floor task on both backbones, and so is L2-T0 on

\pi_{0}

Task	Base	AEGIS $[95\%\ \mathrm{CI}]$	FailBank $[95\%\ \mathrm{CI}]$
L1-T0 Apple	0.368	0.407 [0.109, 0.707]	0.261 [0.004, 0.535]
L1-T1 Lemon	0.368	0.524 [0.299, 0.723]	0.446 [0.250, 0.619]
L1-T2 Mango	0.368	0.448 [0.268, 0.613]	0.786 [0.657, 0.888]
L1-T3 Onion	0.368	0.337 [0.199, 0.459]	0.543 [0.427, 0.649]
L1-T4 Tomato	0.368	0.735 [0.493, 0.926]	0.571 [0.341, 0.754]
L2-T0 Apple	0.368	0.307 [0.243, 0.363]	0.437 [0.374, 0.492]

Appendix

Table 3: Per-task BRS with 95% bootstrap intervals. L2-T1 and L2-T4 were evaluated after the score was defined.

Task	Base $\mathrm{BRS}$	AEGIS $\mathrm{BRS}$ [95% CI]
L1-T2 Mango	0.368	0.245 [0.009, 0.607]
L1-T3 Onion	0.368	0.615 [0.564, 0.663]
L2-T0 Apple	0.368	0.402 [0.350, 0.446]
L2-T1 Lemon	0.368	0.523 [0.438, 0.609]
L2-T2 Mango	0.368	0.515 [0.209, 0.766]
L2-T4 Tomato	0.368	0.387 [0.271, 0.502]

Appendix

Table 4: AEGIS BRS on

\pi_{0}

tasks with defined normalization.

Model	Arena L1 SR	Action interface	Scope decision
$\pi_{0.5}$ -FFT	0.64	Flow matching	Main backbone with complete matched evaluation
$\pi_{0}$ / $\pi_{0}$ -FFT	0.74 / 0.76	Flow matching	Second backbone with complete matched evaluation
$\pi_{0}$ -FAST / -FFT	0.40 / 0.60	FAST tokens	Inapplicable because the chunk-level continuous action loss is absent
OpenVLA	0.60	Discrete autoregressive	Inapplicable to the current loss and guard
OpenVLA-OFT	0.20	Discrete autoregressive	Inapplicable to the current loss and guard
UniVLA	0.42	Autoregressive latent action	Inapplicable to the current loss and guard

Appendix

Table 5: Audit of Arena-finetuned backbone candidates. Inapplicable models are architectural exclusions rather than negative method results.

Evidence	Suite / level	Tasks	Protocol
Main static	static obstacles / L1	T0–T4	base, AEGIS, FailBank , 50 states
State holdout	static obstacles / L1	T2	train 32 states, test 18 unseen states
Zero-shot static	static obstacles / L2	T0–T4	base, AEGIS, 3 FailBank orders
Static rounds	static obstacles / L1	T3	base and R1–R5, 50 states
Dynamic stress test	dynamic obstacles / L1	T3, T4	base and learned policies, 18 states
Dynamic rounds	dynamic obstacles / L2	T0	base and R1–R3, 50 states

Appendix

Table 6: Complete ledger of reported and exploratory task coverage. Base-only checks are not method comparisons.

Radius	Training exposure	Evaluation	Result
Unseen states	L1-T2 training states	held-out L1-T2 states	SR $71.1\%\!\rightarrow\!93.3\%$
Unseen tasks	L1-T2 only	other Level 1 tasks on $\pi_{0.5}$	mean SR and paired test in Table 25
Unseen level	Level 1 only	Level 2 T0–T4	apple/tomato gains, mango/onion preserved, lemon floor

Appendix

Table 7: Three widening generalization radii. Only L1-T2 rollouts enter the self-evolution bank. Split details are given above.

Category	Suite	Cost semantics	Count	Measured evidence	Evaluation status
Safety	Static obstacles	Fall and contact with protected external bodies	10/15	Complete matched three-arm results	Main method comparison
Safety	Dynamic obstacles	Contact with moving bodies and fall	15/15	Base L1-T0 SR 95%, dynamic transfer fails	Negative stress test
Safety	Cautious grasp	Part-level gripper distance to target itself	15/15	Base SR 0/100. Oracle hazard set empty	Floor and unrepresented constraint
Safety	State preservation	Containment predicates	10/15	Base SR 66/100. Zero action rewrites in 196 AEGIS chunks	Incompatible avoidance semantics
Safety	Hazard avoidance	Surface-distance dwell cost near stove/candle and fall	15/15	Initial objects in cost region 42–50/50. VLM misidentifies both preliminary validity cells	Teacher pilot and AEGIS validity gate fail
Distractor	Static distractors	No cost predicate	0/15	BDDL audit	No benchmark cost axis

Appendix

Table 8: Evaluation-scope audit from BDDL predicates and measured probes. The count column gives the number of tasks with a cost predicate followed by the number of tasks inspected. A missing cost predicate provides no benchmark cost axis. It is not a negative method result.

Level	Task	$\pi_{0.5}$ base			$\pi_{0}$ base
		SR	CC	Logged policy cost	SR	CC	Logged policy cost
L1	T0	9.0	16.75	334.9	4.0	15.99	319.7
	T1	0.0	21.55	430.6	2.0	16.68	333.2
	T2	62.0	8.09	161.8	2.0	11.12	222.3
	T3	34.0	15.79	315.6	12.0	20.36	406.8
	T4	0.0	20.67	413.5	0.0	22.40	448.0

Appendix

Table 9: Hazard-Avoidance base-policy headroom. The evaluation includes all 20 base-policy task arms and 50 offsets per arm. CC is the archived official cost; the final column reports the archived policy-induced cost. These columns use different scales and cannot be interpreted as the additive decomposition in Equation 9 . Cost-unit reconciliation is required before comparing them.

Figure 7: Observe-only versus shield-in-loop collection on Level 1 mango. Panel a shows paired episode outcomes, with observe-only outcomes in rows and shield-in-loop outcomes in columns. Panel b groups failure records by paired episode outcome. Panel c counts stored early pre-contact records before outcome-aware admission.

Collection	Train records	Guard records	Guard decision	Apple SR
Observe-only	3,657	300	pass	34.0
Shield in loop	5,099	112	reject	unavailable

Appendix

Table 10: Collection-mode training outcome under the main update recipe.

Update target	Seed 1	Seed 2	Order 3	Mean
Base	8.0	8.0	8.0	8.0
SFTPOS	4.0	10.7	10.7	8.4
SFT0	18.0	14.7	25.3	19.3
SHAM	26.0	16.0	18.0	20.0
FailBank	36.0	31.3	35.3	34.2

Appendix

Table 11: Level 2 apple SR attribution controls over three training orders.

Task	Base	FailBank	SFT0	SHAM	SFTPOS
L1-T1 Lemon	83.0 / 35.87	83.0 / 12.80	84.0 / 5.66	94.0 / 7.75	81.0 / 29.01
L1-T4 Tomato	85.0 / 37.01	90.0 / 31.24	86.0 / 33.51	82.0 / 40.61	81.0 / 51.27

Appendix

Table 12: Level 1 attribution controls for one training order. Each cell reports SR followed by policy-induced CC over 50 offsets under two sampler conditions.

Task	Method	SR	CC	$\mathrm{CC}_{\mathrm{policy}}$
Apple	Base	8.0	67.51	65.89
	AEGIS	7.3	90.73	89.26
	Teacher in loop	0.0	67.02	67.01
Mango	Base	89.3	51.41	33.93
	AEGIS	88.0	28.75	11.15
	Teacher in loop	50.7	107.63	97.50

Appendix

Table 13: Executing the privileged CBF teacher as a shield. Cost cells report CC followed by policy-induced CC.

Method	Updated VLA	Shield	Object geometry	QP/step	s/step
Base policy	no	no	no	no	0.362
AEGIS	no	yes	yes	yes	0.449
FailBank	LoRA	no	no	no	0.346

Appendix

Table 14: Deployment requirements. Privileged object geometry is an offline teacher input for FailBank and is absent at evaluation.

Component	Setting
PaliGemma 2B LoRA	rank 16, $\alpha=16$
300M action-expert LoRA	rank 32, $\alpha=32$
Trainable parameters	LoRA $A/B$ factors only
Optimizer	AdamW, gradient clipping 1.0
Learning-rate schedule	cosine, 20-step warmup, peak $3\!\times\!10^{-5}$
	400-step decay, final $3\!\times\!10^{-6}$

Appendix

Table 15: Policy-update hyperparameters.

Bank	Collection suite	Train records	CBF-triggered / quiet	Validation fold	Used by
Main static two-round	static L1-T2	6,535	2,863 / 3,672	600	main tables, L2, R2 ablations
Pre-filter static pool	static L1-T2	12,343	4,485 / 7,858	600	audit only
State holdout	static L1-T2	4,006	1,799 / 2,207	600	unseen-state evaluation
Static R1–R5	static L1-T2	3,657 / 6,535 / 15,707 / 19,051 / 22,732	–	–	round curve
Dynamic diagnostic R1	dynamic L2-T0	3,192	642 / 2,550	139	dynamic L1 stress test
Dynamic diagnostic R2	dynamic L2-T0	9,898	1,818 / 8,080	266	dynamic round analysis

Appendix

Table 16: Learning-record banks used by the reported experiments. CBF-triggered and quiet counts refer to records with and without a triggered correction in the filtered training fold. They are not counts of corrective and nominal targets.

Figure 8: Training and self-evolution diagnostics. Panels a and b show observe-only constraint activation for static tasks and dynamic rounds. Panel c separates accumulated training records by trigger provenance. Panel d shows held-out full-chunk flow-loss ratios relative to base. The dashed line marks the acceptance limit. Activation is a diagnostic, not a task-success or cost metric.

Bank	Records	Flow-loss ratio	First-action drift	Triggered loss	Decision
Observe-only round 1	3,657	0.972	–	–	accept
Shield-in-loop round 1	5,099	1.137	0.0129	–	reject
Round 2 only	2,878	1.111	0.0053	–	reject
Accumulated, size matched	2,878	1.008	–	–	accept
Accumulated, full	6,535	1.0065	0.00949	–	accept

Appendix

Table 17: Training-guard decisions for the bank-construction study. All rows use 800 update steps and one training seed. A dash denotes an unreported diagnostic, not a zero value.

$\tau_{\mathrm{loss}}$	$\tau_{\mathrm{drift}}=0.01$	0.02	0.05	0.10
1.05	87	89	89	89
1.10	96	100	100	100
1.15	100	108	108	108
1.20	100	109	109	109
1.30	101	110	110	110

Appendix

Table 18: Acceptance counts under alternative guard thresholds.

			L2 Apple		L1 Tomato
Steps	Flow-loss ratio	First-action drift	SR	$\mathrm{CC}_{\mathrm{policy}}$	SR	$\mathrm{CC}_{\mathrm{policy}}$
Base	–	–	8.0	65.89	85.0	37.01
400	1.0042	0.0083	32.0	61.71	88.0	36.65
800	1.0073	0.0096	36.0	65.73	90.0	31.24
1,600	1.0165	0.0108	46.7	56.92	85.0	35.71
2,400	1.0341	0.0121	54.0	43.01	90.0	29.38

Appendix

Table 19: Training duration with the acceptance guard disabled.

Steps	$\lambda_{q}$	Flow-loss ratio	Triggered loss	Decision
800	0.0	1.477	0.185	reject
500	0.0	1.513	0.171	reject
400	0.0	1.580	0.170	reject
800	0.3	1.110	0.199	reject
800	0.5	1.083	0.178	accept
800, $k=10$	0.0	1.017	–	accept

Appendix

Table 20: Held-out guard search and multi-step follow-ups for

\pi_{0}

. The flow-ratio limit is 1.10. Multi-step triggered losses were not reported in the audited summary and are marked unavailable.

Task	Base	$k=1,\lambda_{q}=0.5$	$k=10,\lambda_{q}=0$	$k=50,\lambda_{q}=0$
L1-T0	54.0	58.0	34.0	30.0
L1-T1	66.0	84.0	60.0	62.0
L1-T2	92.0	96.0	94.0	90.0
L1-T3	28.0	26.0	24.0	14.0
L1-T4	56.0	88.0	78.0	78.0
Pooled SR and policy-induced CC	59.2 / 3.78	70.4 / 2.09	58.0 / 2.23	54.8 / 2.94

Appendix

Table 21:

\pi_{0}

multi-step supervision ablation. Values are SR over 50 matched offsets per task. The pooled row reports mean SR followed by mean policy-induced CC.

Perturbation	Bank size	Corrected targets	Affected share, %	$\Delta$ weight mass, %
Default	6,535	709	0.0	0.0
Risk cutoff $\times 0.8$	7,683	457	21.9	$+19.3$
Risk cutoff $\times 1.2$	4,892	1,293	35.6	$-31.0$
Stage boundaries 24 / 12	6,626	744	1.4	$+1.9$
Stage boundaries 36 / 18	6,459	682	1.2	$-1.2$
$\eta$ bins $-0.05$	6,535	709	19.2	$+26.3$

Appendix

Table 22: Record and teacher threshold sensitivity.

Task	Base	Observe-only R1	In-loop R1	Accumulated matched	R2 only	Full accumulated
L2-T0 Apple	8.0 / 65.89	34.0 / 59.97	18.0 / 57.68	41.3 / 48.07	38.7 / 56.59	36.0 / 65.73
L1-T1 Lemon	83.0 / 35.87	85.0 / 24.00	93.0 / 3.24	89.0 / 10.61	87.0 / 8.51	83.0 / 12.80
L1-T4 Tomato	85.0 / 37.01	89.0 / 30.57	87.0 / 30.55	87.0 / 38.53	89.0 / 31.23	90.0 / 31.24

Appendix

Table 23: Non-deployable diagnostics with the acceptance guard disabled. Each cell reports SR followed by policy-induced CC. Bold labels mark rejected variants, not preferred results.

Figure 9: Completion timing and policy-induced CC distributions on Level 2 apple, mango, and onion. Panels a–c show the fraction of all trials completed by each executed control-step count. Panels d–f show the fraction whose policy-induced CC exceeds each threshold. Blue shading spans three training orders and is not a confidence interval.

Task	Base	AEGIS	FailBank	FailBank + AEGIS
L2-T2 Mango	89.3 / 51.41	88.0 / 28.75	94.0 / 36.78	86.0 / 19.78
L2-T3 Onion	81.3 / 16.33	1.3 / 0.27	84.7 / 16.93	2.7 / 0.53

Appendix

Table 24: Stacking AEGIS onto the learned policy. Each cell reports SR followed by

\mathrm{CC}

over 150 matched cells.

Analysis	Comparison	Favorable / adverse / tie	$p$ -value
Breadth	Static L1-T0 SR	7 / 7 / 36	1.0000
	Static L1-T1 SR	12 / 10 / 28	0.8320
	Static L1-T3 SR	17 / 2 / 31	0.0007
	Static L1-T4 SR	11 / 7 / 32	0.4810
	Four unseen L1, SR $82.5\%\to 88.7\%$	51 / 45 / 104	0.61
	Four unseen tasks, legacy policy-induced CC	71 / 53 / 76	0.1265

Appendix

Table 25: Detailed paired tests for breadth and learning-signal ablations. Favorable means higher SR or lower policy-induced CC. Individual L1 task rows and legacy cost rows use the earlier no-curriculum configuration. Pooled SR rows use training-order means. Unreported tie counts are marked unavailable.

Backbone	Set	Sign-test $p$	$\Delta$ SR	95% CI	$\Delta\mathrm{CC}_{\mathrm{policy}}$	95% CI
$\pi_{0.5}$	L1 unseen four	0.61	+6.2	[2.0, 10.5]	-8.39	[-15.21, -1.97]
	L1 all five	0.099	+8.3	[4.5, 12.3]	-17.24	[-24.22, -10.61]
	L2 all five	$<0.0001$	+8.8	[6.2, 11.6]	-16.78	[-24.16, -9.40]
	All ten	$<0.0001$	+8.6	[6.2, 10.9]	-17.01	[-22.19, -12.00]
$\pi_{0}$	L1 unseen four	0.014	+11.8	[5.0, 18.8]	-1.59	[-3.95, 0.53]
	L1 all five	0.024	+9.6	[3.8, 15.6]	-1.08	[-2.97, 0.62]

Appendix

Table 26: Effect sizes with 95% bootstrap intervals. SR changes are percentage points. Intervals resample initial states only and exclude variation across training orders. The non-floor L2 row is post hoc.

Task	Training	SR higher / lower / equal	SR $p$ -value	Activation higher / lower / equal	Activation $p$ -value
	order
Apple	1	28 / 2 / 20	$8.68\!\times\!10^{-7}$	32 / 18 / 0	0.0649
	2	24 / 3 / 23	$4.92\!\times\!10^{-5}$	22 / 26 / 2	0.6655
	3	29 / 3 / 18	$2.56\!\times\!10^{-6}$	30 / 20 / 0	0.2026
Mango	1	8 / 4 / 38	0.3877	14 / 36 / 0	0.0026
	2	10 / 3 / 37	0.0923	14 / 36 / 0	0.0026

Appendix

Table 27: Repeated sampler conditions are averaged, not counted as independent samples. SR is success rate. Activation is the observe-only constraint-flag rate.

Task	Arm	SR	CC	$\mathrm{CC}_{\mathrm{policy}}$	SR test relative to base
L2-T1 Lemon	Base	0.0	158.14	158.13	–
	AEGIS	0.0	141.51	141.51	–
	FailBank mean	0.0	110.14	110.12	floor
L2-T4 Tomato	Base	74.0	67.87	53.07	–
	AEGIS	28.7	26.93	21.20	$p<0.0001$
	FailBank order 1	84.0	59.05	42.39	$p=0.0872$

Appendix

Table 28: Pre-registered Level 2 completion experiment. Each training-order comparison uses 50 matched offsets under three sampler conditions. FailBank mean is the average of the three orders.

Task	Cost	FailBank -1	FailBank -2	FailBank -3	AEGIS
Apple	$\mathrm{CC}_{\mathrm{policy}}$	25/25/0, 1.000	25/25/0, 1.000	27/23/0, 0.672	19/27/4, 0.302
	CC	22/28/0, 0.480	23/27/0, 0.672	26/24/0, 0.888	18/27/5, 0.230
Mango	$\mathrm{CC}_{\mathrm{policy}}$	28/20/2, 0.312	34/16/0, 0.015	28/20/2, 0.312	42/8/0, $1.2{\times}10^{-6}$
	CC	29/19/2, 0.193	35/15/0, 0.0066	29/21/0, 0.322	42/8/0, $1.2{\times}10^{-6}$
Onion	$\mathrm{CC}_{\mathrm{policy}}$	1/0/49, 1.000	1/1/48, 1.000	1/0/49, 1.000	1/0/49, 1.000
	CC	9/13/28, 0.524	11/11/28, 1.000	10/11/29, 1.000	49/0/1, $3.6{\times}10^{-15}$

Appendix

Table 29: Offset-paired Level 2 cost tests.

\mathrm{CC}_{\mathrm{policy}}

removes the benchmark’s initial-state component. CC follows the benchmark definition.

Explore similar work

Failing Forward: Adaptive Failure-Informed Learning for Vision-Language-Action Models
May 8, 2026Meng Zheng, Samhita Marri, Anwesa Choudhuri +6Action Generation Negative Results
United Imaging Intelligence, Boston, MA, USA · University of Illinois Urbana-Champaign, Urbana, IL, USA
FailPatch: Failure Residual Patching for Vision-Language-Action Models
Sep 28, 2026Peng Yu, Jiacheng Wang, Ziheng Zhang +6Photometric Supervision Robot State
Xi’an Jiaotong University · Dexmal · Nanjing University
RePO-VLA: Recovery-Driven Policy Optimization for Vision-Language-Action Models
May 10, 2026Weijia Liufu, Xiaoyu Guo, Ruiyi Chen +16Frictive Policy Optimization Imitation
1Sun Yat-sen University · 2South China University of Technology · 3Peng Cheng Laboratory +3

May 8, 2026, cs.ROJ/K move · Enter open · S save

Failing Forward: Adaptive Failure-Informed Learning for Vision-Language-Action Models

Meng Zheng, Samhita Marri, Anwesa Choudhuri, Benjamin Planche+5

United Imaging Intelligence, Boston, MA, USA · University of Illinois Urbana-Champaign, Urbana, IL, USA

Action Generation Negative Results

Open paper AskPDF arXiv

Vision-language-action (VLA) models provide a promising paradigm for scalable robotic manipulation, yet their reliance on success-only behavioral cloning leaves them brittle; lacking corrective training signals, minor execution errors rapidly compound into unrecoverable, out-of-distribution failures. To address this limitation, we propose Adaptive Failure-Informed Learning (AFIL), an end-to-end framework that leverages failure trajectories as adaptive negative guidance for diffusion- and flow-based VLA policies. AFIL uses a pretrained VLA to generate failure rollouts online, avoiding the need for handcrafted failure-mode design or human-in-the-loop recovery. It then jointly trains Dual Action Generators (DAGs) for successful and failed behaviors while sharing a common vision-language backbone, enabling efficient failure-aware policy learning with limited parameter overhead. During sampling, the failure generator adaptively steers action generation away from failure-prone regions and toward more reliable success modes, with guidance strength determined by the per-diffusion-step distance between success and failure distributions. Experiments across in-domain and out-of-domain robotic manipulation tasks, covering both short- and long-horizon settings, show that AFIL consistently improves task success rates and robustness over existing VLA baselines, demonstrating its effectiveness, efficiency, and generality.

Abstract

Explore similar work

Failing Forward: Adaptive Failure-Informed Learning for Vision-Language-Action Models

FailPatch: Failure Residual Patching for Vision-Language-Action Models

RePO-VLA: Recovery-Driven Policy Optimization for Vision-Language-Action Models