Natural State-Prediction Accuracy can Hide Weak Controlled Responsiveness in VLA Readouts

Hyungjoon Kim; Wonbin Son; Mi Young Lee; Jun Young Lee; Seungmin Rho

cs.ROSep 28, 2026

Natural State-Prediction Accuracy can Hide Weak Controlled Responsiveness in VLA Readouts

Authors: Hyungjoon Kim, Wonbin Son, Mi Young Lee, Jun Young Lee, Seungmin Rho

Organizations: Changwon National University · Chung-Ang University

Abstract

Accurately decoding object states from the internal representations of vision-language-action (VLA) models does not establish that the predictions respond faithfully to changes in the target physical state. In natural observations, object state, robot configuration, occlusion, and task progress vary together, allowing contextual cues to contribute to prediction. In this paper, we introduce an evaluation framework that separates prediction accuracy, target-state responsiveness, and context stability using physically validated observations that cross target coordinates with robot contexts. We demonstrate that high natural-trajectory accuracy can coexist with weak controlled target-state responsiveness in fixed representation-readout pairs. Comparisons and interventions involving representations, readouts, and training data show that the three properties provide distinct diagnostic information. Furthermore, adding responsiveness and context sensitivity to a failure predictor based on initial state error and physical variables reduces policy-failure prediction error on new initializations relative to the specified baseline while same-observation controlled MAE is also informative. These findings motivate evaluating target-state responsiveness and context stability alongside natural prediction accuracy, and examining their relationship to actual policy behavior and task outcomes.

Figures & tables

Figure 1: Evaluation framework. (a) Target state and robot context co-vary in natural observations. (b) Crossed comparisons vary state at fixed context (orange row) and context at fixed state (blue column). (c) The same fixed representation–readout pair is evaluated for accuracy, responsiveness, and stability. A separate fixed policy receives observations, and its outcomes are related to the diagnostics (dashed connection). Readout predictions are not policy inputs. Scenes, grids, and small plots are schematic illustrations.

Evaluation	Data and independent units	Purpose
Primary confirmation	24 train / 8 development / 12 confirmation initializations per task; 4 contexts $\times$ 6 coordinates per scene	Natural accuracy, controlled response, matched training
Representation/model extension	32 new evaluation initializations per task; radial/lateral context axes	Fixed-candidate replication and comparison
Free object	32 mug-position initializations; 64 support follow-up initializations	Physical scope and estimability
Failure prediction	96 development / 128 evaluation initializations; 2 normal action-noise rollouts each	Additional pre-rollout outcome information

Table 1: Evaluation design. Behavioral experiments use a task-adapted, frozen SmolVLA policy in LIBERO ( Liu et al., 2023 ) , separately from the Meta-World representation analysis. Appearance interventions, stove evaluations, and development controls use the distinct sample counts reported with their results.

Task	Model	Nat. MAE	Ctrl. MAE	$G$ [95% CI]	$E_{G}$	$R$
Drawer	SmolVLA	5.554	74.184	0.1806 [0.1653, 0.1978]	0.8194	16.499
	OpenVLA	3.892	75.884	0.1525 [0.1387, 0.1662]	0.8475	9.797
Faucet	SmolVLA	0.0763	0.6698	0.0211 [0.0035, 0.0388]	0.9789	0.1285
	OpenVLA	0.0570	0.6475	0.0512 [0.0292, 0.0717]	0.9488	0.1008

Table 2: Primary confirmation with full-natural readouts. Values are equal-weight means over 12 scenes per task. MAE and

R

use mm for Drawer and rad for Faucet;

G

and

E_{G}

are dimensionless. Brackets are individual 95% scene-bootstrap intervals. Unit mean gain does not imply accuracy at every intermediate state. Appendix A.7 reports intervals for all metrics.

Figure 2: Confirmatory predictions and responses. Left: all 12 scenes

\times

4 context curves per combination (faint), mean predictions (bold), and identity lines (dashed). Right: scene-level endpoint gains and post-hoc all-coordinate slopes; large markers and bars show means and individual 95% intervals. Both summaries are below unit response. Mean gain does not characterize intermediate non-monotonicity.

Model	Drawer $G$ : radial / lateral	Faucet $G$ : radial / lateral
SmolVLA	0.2058 / 0.2060	0.0097 / 0.0122
OpenVLA	0.1692 / 0.1661	0.0429 / 0.0394
$\pi_{0}$ base	0.1094 / 0.1086	$-0.0106$ / $-0.0011$
GR00T N1.6	0.2413 / 0.2426	0.0428 / 0.0435

Table 3: Mean endpoint gain of fixed final-mean-ridge candidates on 32 new initializations per task. Radial and lateral axes share initializations and use local

\pm 10

mm hand-position changes relative to the robot. Model-native input paths differ, so this is not a controlled ranking of model quality. Appendix B provides MAE and candidate details.

Evaluation	Initializations	Result
6 cm position variation	32	Natural MAE 128.314 mm; controlled MAE 39.765 mm; $G=0.140$ . Natural error is already large.
Support-conditioned follow-up	64	43 physically valid; 5 support-valid; intersection 0. Below the minimum of 32 for estimation.

Table 4: Free-object results and estimability. The first evaluation passes its physical and visual criteria. The support follow-up is non-estimable, not a zero-gain or model-response failure. The target is one position component, not full 6-DoF state.

Figure 3: Matched controlled minus matched natural training. Points and bars show means and individual 95% paired scene-bootstrap intervals over 12 confirmation scenes. Negative values indicate lower error or sensitivity. For display only, MAE and

R

are divided by

S=0.16

m for Drawer and

S=1.4

rad for Faucet; metric definitions are unchanged.

Model	Information added to $M_{0}$	Brier	Log loss
$M_{0}$	None	0.181515	0.543533
$M_{C}$	Controlled MAE	0.175654	0.528602
$M_{GR}$	$E_{G},R$	0.171381	0.519075
$M_{CGR}$	Controlled MAE, $E_{G},R$	0.169677	0.514777

Table 5: Failure-prediction losses on the same 128 evaluation initializations; lower is better. Two normal-noise rollouts are grouped within each equally weighted initialization.

M_{0}/M_{GR}

form the original fixed comparison;

M_{C}/M_{CGR}

are post-hoc additions. This evaluates failure prediction, not an improvement to policy success.

Figure 4: Mean loss differences and individual 95% intervals from 20,000 paired initialization-bootstrap draws. Negative differences favor the first model. Original and post-hoc comparisons are distinguished. Intervals describe evaluation uncertainty conditional on the development fits.

Appendix figures & tables30 assets

Supplementary material from the paper’s appendix.

Appendix

Condition	Artifact name	Training observations
Full natural	natural_full	All 24 natural training trajectories
Matched natural	natural36	Six scenes, six observations each; 36 total
Matched controlled	exact_control36	The same six scenes and 36 actual coordinates

Appendix

Table S1: Primary training conditions. Implementation names are included solely to map stored artifacts to scientific conditions.

Task	Training	Development	Confirmation
Drawer	108200–108223	108300–108307	108400–108411
Faucet	110200–110223	110300–110307	110400–110411

Appendix

Table S2: Primary simulator initialization identifiers.

Model	Feature	Dimension/pooling
SmolVLA	Final prefix image tokens	Four spatial means concatenated; 3,840 primary
OpenVLA	Final-layer 256 causal image states	Mean; 4,096 primary
$\pi_{0}$ base	Final native image/prefix features	Mean or adaptive $2\times 2$ ; no numerical robot state in this feature path
GR00T N1.6	Final native visual tokens	Native $9\times 9$ grid; mean or adaptive $2\times 2$
SigLIP2	Vision pooler output	768; revision-pinned processor
VC-1	Normalized CLS	768; official resize/crop and ImageNet normalization

Appendix

Table S3: Frozen feature paths. Model-native token grids and preprocessing are retained.

Task	Metric	SmolVLA	OpenVLA
Drawer	Natural MAE	5.554 [4.921, 6.328]	3.892 [3.357, 4.477]
	Controlled MAE	74.184 [72.277, 75.993]	75.884 [74.810, 76.974]
	$G$	0.1806 [0.1653, 0.1978]	0.1525 [0.1387, 0.1662]
	$E_{G}$	0.8194 [0.8022, 0.8347]	0.8475 [0.8338, 0.8613]
	$R$	16.499 [13.189, 20.415]	9.797 [7.576, 12.035]
Faucet	Natural MAE	0.0763 [0.0596, 0.0940]	0.0570 [0.0446, 0.0707]

Appendix

Table S4: Complete primary metrics: mean [individual 95% scene-bootstrap interval],

n=12

per task. MAE/

R

are mm for Drawer and rad for Faucet;

G,E_{G}

are dimensionless.

Figure S1: All primary scene-level context sensitivities

R

, with means and individual 95% intervals. Each point is one initialization; contexts and observations are not independent sample units.

Task	Model	Matched natural MAE	Matched controlled MAE	$G$
Drawer	SmolVLA	4.155 mm	49.369 mm	0.1217
	OpenVLA	3.242 mm	49.317 mm	0.1451
Faucet	SmolVLA	0.1025 rad	0.5976 rad	0.0193
	OpenVLA	0.0755 rad	0.5873 rad	0.0442

Appendix

Table S5: Actual-coordinate-matched development evaluation.

Task	Model	Readout	Nat. MAE	Ctrl. MAE	$G$	$E_{G}$	$R$
Drawer	SmolVLA	Linear	4.846	76.295	0.1785	0.8215	15.514
		RBF	6.068	75.934	0.1696	0.8304	15.316
		MLP	3.650	78.502	0.1506	0.8494	13.176
	OpenVLA	Linear	3.606	76.217	0.1487	0.8513	10.534
		RBF	4.590	77.307	0.1262	0.8738	9.752
		MLP	2.408	78.310	0.1337	0.8663	8.822

Appendix

Table S6: Representative full-natural development results. MAE/

R

are mm for Drawer and rad for Faucet. RBF uses fixed

\sigma=\tilde{d}

\lambda=0.001

; MLP averages three seed-specific metrics.

Task	Model	Natural MAE	Controlled MAE	$G$
Drawer	SmolVLA	7.303	72.420 / 74.559	0.2058 / 0.2060
	OpenVLA	4.317	76.258 / 76.555	0.1692 / 0.1661
	$\pi_{0}$ base	6.689	81.547 / 82.161	0.1094 / 0.1086
	GR00T N1.6	3.711	69.011 / 68.812	0.2413 / 0.2426
Faucet	SmolVLA	0.0698	0.6867 / 0.6792	0.0097 / 0.0122
	OpenVLA	0.0484	0.6789 / 0.6778	0.0429 / 0.0394

Appendix

Table S7: Fixed final-mean-ridge results on 32 new initializations per task. MAEs are mm for Drawer and rad for Faucet. Paired entries denote radial/lateral axes.

Task	Model	Endpoint $G$	All-coordinate $B$
Drawer	SmolVLA	0.1806 [0.1653, 0.1978]	0.1618 [0.1477, 0.1762]
	OpenVLA	0.1525 [0.1387, 0.1662]	0.1561 [0.1427, 0.1691]
Faucet	SmolVLA	0.0211 [0.0035, 0.0388]	0.0262 [0.0141, 0.0373]
	OpenVLA	0.0512 [0.0292, 0.0717]	0.0550 [0.0394, 0.0702]

Appendix

Table S8: Endpoint gains and all-coordinate slopes: means [individual 95% intervals].

Task	Model	$R$ increases	$R$ decreases
Drawer	SmolVLA	12	0
	OpenVLA	10	2
Faucet	SmolVLA	11	1
	OpenVLA	9	3

Appendix

Table S9: Scene-level context-sensitivity changes under matched controlled replacement (

n=12

per task).

Design	Composition	Purpose
Joint-A	36 matched natural + 36 matched controlled = 72	Combine observations of the same scenes/coordinates
Joint-B	Full natural + 36 controlled; 688 Drawer / 515 Faucet	Retain all original natural training data
Fixed budget	36 total; controlled fractions $1/3,1/2,2/3$ , each with two complementary assignments	Compare mixtures at fixed sample count

Appendix

Table S10: Joint-training designs. All fit a single readout to both observation types.

Task	Model	Pooling	Natural MAE	Controlled MAE	$G$
Drawer	SmolVLA	Full mean	7.501	72.656	0.200
		Object ROI	10.038	42.553	0.521
		Equal-area control	6.736	85.426	0.108
	OpenVLA	Full mean	3.606	76.217	0.149
		Object ROI	5.708	41.143	0.514
		Equal-area control	3.586	70.235	0.243

Appendix

Table S11: Representative natural-full region-pooling controls. MAE units are mm for Drawer and rad for Faucet. Blank task/model cells inherit the preceding entry.

Evaluation	Roots	Natural MAE (mm)	Controlled MAE (mm)	$G$
Development	16	121.767	42.293	0.167
New roots	32	128.314	39.765	0.140

Appendix

Table S12: Free-object results with 6 cm state variation. Both evaluations pass their physical and visual criteria. Natural error is already large.

Gate	Roots
Physically valid	43
Prespecified support-valid	5
Physical and support-valid	0
Required minimum for response estimation	32

Appendix

Table S13: Free-object follow-up eligibility. No root satisfies both required gates.

Condition	Mean $\|\Delta\hat{q}\|$ (mm)	First translation-command shift
Robot-low	4.482	0.03613
Robot-high	15.506	0.04550

Appendix

Table S14: Initial changes relative to sham under identical physical state. Translation shifts are normalized command

L_{2}

differences, not distances in physical coordinate units.

Input	Definition	Scaling
$E_{0}$	Absolute prediction error at the normal initial observation	Divide by 0.14 m
$D_{q}$	Initial coordinate distance to success boundary $-0.14$ m	Divide by 0.14 m
$D_{\mathrm{eef}}$	Euclidean distance from end effector to cabinet origin	m
$D_{\mathrm{robot}}$	Standardized RMS difference of seven arm joints from policy-training state mean	Policy-training standard deviation; dimensionless
$E_{G}$	Mean context-wise endpoint gain error from six controlled snapshots	Dimensionless
$R$	Prediction range over three contexts, averaged over two coordinates	Divide by 0.14 m

Appendix

Table S15: Failure-predictor inputs, all measured before rollout.

Loss	$M_{0}$	$M_{GR}$	Difference	Individual 95% interval
Brier	0.181515	0.171381	$-0.010134$	[ $-0.017898$ , $-0.002478$ ]
Log loss	0.543533	0.519075	$-0.024458$	[ $-0.043453$ , $-0.005586$ ]

Appendix

Table S16: Original fixed failure-prediction comparison. Intervals condition on development fitting.

Successes among two rollouts	Roots	Mean whole-trajectory MAE (mm)
0/2	28	55.075
1/2	25	35.295
2/2	75	6.701

Appendix

Table S17: Post-hoc relationship between whole-trajectory error and success, grouping two normal rollouts per evaluation root.

Task	Original	Sham	Robot-low	Robot-high	BG-low	BG-high
Drawer	41	41	41	36	42	41
Stove	60	60	61	60	60	60

Appendix

Table S18: Appearance-intervention successes out of 64 rollouts per task/condition (32 roots, two noises).

Task	Central	$-0.12$ rad	$+0.12$ rad
Drawer	8	3	1
Stove	8	8	8

Appendix

Table S19: Initial development-pilot successes out of eight under different robot contexts.

Model	Added to $M_{0}$	Brier	Log loss
$M_{0}$	None	0.181515	0.543533
$M_{C}$	Controlled MAE	0.175654	0.528602
$M_{GR}$	$E_{G},R$	0.171381	0.519075
$M_{CGR}$	Controlled MAE, $E_{G},R$	0.169677	0.514777

Appendix

Table S20: All failure-prediction models on 128 evaluation roots.

Comparison	Brier difference	Log-loss difference
$M_{CGR}-M_{C}$	$-0.005977$ [ $-0.012421$ ,0.000381]	$-0.013824$ [ $-0.030022$ ,0.002338]
$M_{GR}-M_{C}$	$-0.004273$ [ $-0.010564$ ,0.001869]	$-0.009527$ [ $-0.025286$ ,0.006211]
$M_{C}-M_{0}$	$-0.005861$ [ $-0.009132$ , $-0.002744$ ]	$-0.014932$ [ $-0.022798$ , $-0.007270$ ]
$M_{GR}-M_{0}$	$-0.010134$ [ $-0.017898$ , $-0.002478$ ]	$-0.024458$ [ $-0.043453$ , $-0.005586$ ]

Appendix

Table S21: Paired loss differences [individual 95% intervals], sharing 20,000 root-resampling indices.

Task	Roots	Planned	Success	Timeout	Unknown
Mug to plate	32	64	13	51	0
Book to compartment	32	64	2	57	5
Middle book to shelf	32	64	1	48	15
Right book to shelf	32	64	11	44	9
Descriptive total	128	256	27	200	29

Appendix

Table S22: Exploratory natural rollouts of fixed pick-and-place policies. Numerical errors with unknown outcomes are not imputed as failures.

Subset	Drawer	Faucet
Main controlled grid	32/32	32/32
Natural- $q$ -matched grid	0/32	18/32
Target- $q$ -support-restricted main grid	0/32	32/32
Support-restricted natural-matched grid	0/32	0/32

Appendix

Table S23: Eligible roots for each analysis, per context axis.

Task	Model	Sham $G$	Robot-high $G$
Drawer	SmolVLA	0.2124	0.1730
	OpenVLA	0.1608	0.1126
	$\pi_{0}$	0.0992	$-0.1009$
	GR00T	0.2388	0.2326
Faucet	SmolVLA	0.0135	0.0042
	OpenVLA	0.0345	$-0.0084$

Appendix

Table S24: Appearance effects on endpoint gain under a common mask.

Encoder	Drawer $G$	Faucet $G$
SigLIP2	0.3357	0.0794
VC-1	0.1979	0.0503
R3M	0.3445	0.1391
DINOv3	0.1837	0.1611

Appendix

Table S25: Native-context gains for non-VLA controls on the common evaluation bank.

Task	Natural MAE	Controlled MAE	$G$	$E_{G}$	$R$
Drawer	3.947	77.460	0.1837	0.8163	7.158
Faucet	0.0414	0.5532	0.1611	0.8389	0.0937

Appendix

Table S26: DINOv3 native-context metrics. MAE/

R

are mm for Drawer and rad for Faucet.

Task	Appearance	Natural MAE	Controlled MAE	$G$	$R$
Drawer	Original/sham	3.947	77.460	0.1837	7.158
	Robot-low	12.127	69.297	0.2036	10.336
	Robot-high	17.906	55.802	0.2587	14.443
	Background-low	6.396	79.937	0.1957	8.143
	Background-high	8.431	80.598	0.2106	6.352
Faucet	Original/sham	0.0414	0.5532	0.1611	0.0937

Appendix

Table S27: DINOv3 appearance-condition metrics, with mm for Drawer MAE/

R

and rad for Faucet MAE/

R

Explanation	Available evidence	Remaining question
Physical/rendering errors	Actual coordinate, robot, visibility, repeated-render and target-pixel checks	Does not establish natural joint-distribution membership
Different target-coordinate ranges	Actual-coordinate matching and separate faucet state-value matching	Does not match the entire context distribution
Variation confined to artificial images	Prediction differences in natural-only near-equal- $q$ pairs	Observational matching does not isolate a causal context factor
One VLA or readout fails	Multiple VLAs, feature paths, pooling, nonlinear readouts, non-VLA controls	Models may share sensitivity to distribution shift
Novel state–context combinations are jointly OOD	Limited support-aware follow-ups	Joint-distribution OOD remains possible

Appendix

Table S28: Alternative explanations and the scope of the available controls.

Panel of Figure 39 — Figure S2: Controlled observations at fixed robot contexts. Each row contains six coordinate settings from one development scene: Drawer seed 108300, robot-context frame 6; Faucet seed 110300, robot-context frame 12. Coordinates increase from left to right, follow the simulator’s native sign convention, and are rounded for display. A fixed $180^{\circ}$ orientation correction presents the rendered images upright. These are illustrative examples; quantitative validation is reported in Appendices A and D.

Explore similar work

Aug 4, 2026cs.RO

How Should Vision-Language-Action Models Use Proprioceptive State?

Recent Vision-Language-Action (VLA) models almost universally take robot proprioceptive state as input, yet wire it in incompatible ways -- serialized into text prompts, projected into the vision-language prefix, or fed directly to the action expert -- and almost always as a single current frame. Three questions remain open: (1) whether, and on which tasks, current state actually improves closed-loop control; (2) how much state history helps, and whether its benefit reflects genuine temporal variation rather than added conditioning capacity; and (3) where state should enter the model -- the vision-language backbone or the action-generation module. We answer these questions through controlled experiments on a flow-matching VLA, fixing the backbone, training data, action representation, and evaluation protocol throughout. We implement five representative interfaces -- discrete state prompt, VLM prefix, action prefix, state expert, and feature modulation -- under matched implementation details, and evaluate them on 45 atomic tasks spanning three task families plus 20 composite tasks; we then sweep the state-history length from 1 to 96 frames to examine how historical state information affects model performance. The experiments yield systematic answers to all three questions, distilled into testable design principles for state-aware VLAs.

Yiren Zhao, Ziyang Chen, Ziyang Rao +5

1The Hong Kong University of Science and Technology (Guangzhou) · 3AI2 Robotics X-Lab · 2Multimedia Laboratory (MMLab), The Chinese University of Hong Kong

Proprioception

Ask PDF

May 31, 2026cs.RO

Beyond Task Success: Behavioral and Representational Diagnostics for WAM and VLA

Vision-language-action (VLA) policies and World-Action Models (WAM) represent two increasingly important paradigms for robotic manipulation. However, it remains unclear whether future prediction in WAMs leads to behaviorally meaningful improvements beyond final task success. In this paper, we ask whether WAMs merely add future prediction, or whether they change robot behavior and internal representations in ways that are actionable for control. We introduce a model-agnostic diagnostic framework that compares WAMs and VLAs through two complementary lenses: behavioral rollout analysis and sparse-autoencoder-based feature analysis. The behavioral protocol measures action dynamics consistency, target-object progress, distractor disturbance, and runtime cost. The feature-space protocol characterizes internal representations as memorized, reactive, or predictive, revealing whether models encode future-oriented structure. Across LIBERO and RoboTwin2.0, we evaluate 7 policies spanning direct VLAs and joint, sequential, and auxiliary WAMs. Our results show that success alone hides key differences: WAMs often improve object-level behavior and target selectivity, but their gains depend on architecture and incur higher inference cost. Sequential WAMs show the clearest predictive structure, while auxiliary and joint WAMs respectively compress or entangle future information. These findings suggest future directions for WAMs design to preserve behaviorally actionable future representations for efficient manipulation.

Hung Mai, Bin Zhu, Tuan Do

National Economics University, Vietnam · N2TP Technology · Singapore Management University +1

World Models

Ask PDF

Aug 13, 2026cs.RO

Decoding Task Progress from VLA Representations

Vision-language-action models (VLAs) are moving rapidly towards deployment as general-purpose manipulation policies, but we currently lack basic tools for understanding what these models represent internally or for monitoring them at runtime. Leveraging ideas from mechanistic interpretability, we probe the residual stream of

π_{0.5}

and find that task progress, the normalized time remaining in a trajectory, is linearly readable from the activations. We find that this signal is present in the pretrained PaliGemma backbone prior to training on any robot-specific data. A single linear probe generalizes to unseen tasks and varies under language counterfactuals when trained on multi-prompt data, but does not enable meaningful steering of the policy. These properties make the signal directly useful for instrumenting deployed VLAs. We use the probe as a simple label-free OOD detector, which detects stalled task progress, and find it competitive with state-of-the-art methods. Our results suggest that VLAs have rich, linearly readable internal representations of semantic quantities like task progress, and that learning to read these signals offers a lightweight, interpretable path toward monitoring deployed visuomotor policies.

Atiksh Bhardwaj, Edward Weiyi Duan, Prithwish Dan +2

Cornell University

Visuomotor Policy Mechanistic Interpretability Decoding

Ask PDF