StableVLA: Towards Robust Vision-Language-Action Models without Extra Data

Abstract: It is infeasible to encompass all possible disturbances within the training dataset. This raises a critical question regarding the robustness of Vision-Language-Action (VLA) models when encountering unseen real-world visual disturbances, particularly under imperfect visual conditions. In this work, we conduct a systematic study based on recent state-of-the-art VLA models and reveal a significant performance drop when visual disturbances absent from the training data are introduced. To mitigate this issue, we propose a lightweight adapter module grounded in information theory, termed the Information Bottleneck Adapter (IB-Adapter), which selectively filters potential noise from visual inputs. Without requiring any extra data or augmentation strategies, IB-Adapter consistently improves over the baseline by an average of 30%, while adding fewer than 10M parameters, demonstrating notable efficiency and effectiveness. Furthermore, even with a 14x smaller backbone (0.5B parameters) and no pre-training on the Open X-Embodiment dataset, our model StableVLA achieves robustness competitive with 7B-scale state-of-the-art VLAs. With negligible parameter overhead (<10M), our approach maintains accuracy on long-horizon tasks and surpasses OpenPi under both synthetic and physical visual corruptions.

• The paper introduces StableVLA, a lightweight fused information-bottleneck adapter that uses covariance-based sigmoid gating to suppress noisy visual channels while preserving manipulation-relevant detail without extra data or corruption training.
• StableVLA improves over VLA-Adapter by 40.2%–139.6% relatively across four LIBERO suites at the harshest corruption level, reaching 82.0% on LIBERO-Spatial, 70.2% on LIBERO-Object, 71.9% on LIBERO-Goal, and 45.3% on LIBERO-Long.
• Real-robot tests show stronger resilience to noise, blur, oil, and plastic obstructions, including a 17.5-point average drop on pick-and-place versus 49.2 points for VLA-Adapter, with fewer than 10M additional parameters.

Motivation and problem statement

Vision-Language-Action (VLA) models are typically evaluated under controlled, idealized visual conditions, yet real-world deployment inevitably involves sensor noise, motion blur, weather effects, and lens obstructions. The paper's central empirical finding is that this gap is severe: VLA-Adapter, which attains a 96% success rate on clean LIBERO data, loses nearly 50 percentage points on average when synthetic natural corruptions are injected, and can collapse to 0% success under severe blur. The authors show this fragility is not idiosyncratic to one architecture—OpenVLA (7B), OpenVLA-OFT (7B), and OpenPi–0.5 (3B) all exhibit consistent degradation in simulation and on physical robots.

The paper argues that prevailing data-centric remedies (corruption augmentation, domain randomization) have two structural limitations: simulating the combinatorial space of real-world corruptions is computationally prohibitive, and augmentation tends to induce memorization of specific noise patterns rather than learning of invariant features. This motivates the alternative question the paper pursues: whether intrinsic robustness can be obtained through architectural design alone, without extra data or corruption-aware training.

Diagnosis: the projector as the vulnerability locus

Through feature-consistency analysis across VLA processing stages, the authors localize much of the feature degradation to the projector module bridging the frozen vision encoder and the LLM backbone. Because the vision encoder is typically frozen to preserve semantic priors, input-level noise propagates through it unchanged, and standard MLP projectors—which process channels independently and act as all-pass filters maximizing $I(\mathbf{X}_v; \mathbf{Z})$ indiscriminately—offer no mechanism to suppress task-irrelevant disturbances before they reach the policy model.

Method: IB-Adapter and Fused IB-Adapter

The proposed solution reformulates modality alignment as an Information Bottleneck (IB) problem, minimizing $$I(\mathbf{X}_v; \mathbf{Z}) - \beta I(\mathbf{Z}; \mathbf{S})$$, where $\mathbf{S}$ denotes clean task-relevant semantics. Building on prior work showing self-attention is equivalent to iterative IB optimization under Gaussian assumptions, the authors derive (in an appendix proposition) that channel-wise IB updates reduce to a channel-attention form $$\mathbf{Z} = \mathbf{V}\sigma(\beta\mathbf{Q}^\top\mathbf{K})$$, where $\sigma$ is Softmax under a categorical latent structure or Sigmoid under an independent Bernoulli latent structure.

The resulting IB-Adapter has three components:

• Subspace covariance modeling: multi-head Gram matrices $\mathbf{G}_h = \mathbf{Q}_h^\top \mathbf{K}_h$ computed over channel correlations across spatial tokens, with keys defined by identity mapping to preserve the intrinsic geometric manifold of visual tokens.
• Sigmoid-based gating: a learnable temperature-scaled sigmoid over the Gram matrix, allowing independent per-channel suppression. Channels carrying uncorrelated sensor noise exhibit low covariance with semantic channels and are gated toward zero without competing with robust channels—an explicit contrast to Softmax's forced inter-channel competition.
• Non-linear feature transformation: a two-layer GELU MLP producing value tokens modulated by the gate.

Because pure covariance filtering can attenuate high-frequency details needed for precise manipulation, the final Fused IB-Adapter is dual-pathway:

$$\mathbf{Z} = \text{MLP}(\mathbf{X}) + \tanh(\lambda)\cdot\text{IB-Adapter}(\mathbf{X})$$

combining a high-fidelity MLP path with a denoising IB path. A Stochastic Pathway Dropout (SPD) rate calibrates the balance per task suite: near-zero dropout for precision-critical tasks (LIBERO-Long), moderate dropout ($p_{\text{drop}} \approx 0.3$) for tasks requiring semantic consistency (CALVIN, LIBERO-Object). The total parameter overhead is fewer than 10M parameters.

Benchmark results

Evaluations use the ImageNet-C corruption protocol (19 corruption types at severity levels 3–5) applied zero-shot during policy rollouts on LIBERO and CALVIN, with no corruption exposure or specialized robustness training. Key findings:

Method	Params	Pretraining	LIBERO-Spatial S5	LIBERO-Object S5	LIBERO-Goal S5	LIBERO-Long S5
OpenVLA	7B	OpenX	14.7	2.7	16.3	7.0
OpenVLA-OFT	7B	OpenX	72.1	52.8	70.3	40.3
OpenPi–0.5	3B	OpenX + web	62.4	76.4	64.2	47.7
VLA-Adapter	0.5B	none	58.5	29.3	47.3	26.2
StableVLA	0.5B	none	82.0	70.2	71.9	45.3

StableVLA improves over its direct baseline VLA-Adapter by 40.2% to 139.6% relative across the four LIBERO suites at severity level 5, and completes more CALVIN subtasks at every corruption level (4.17 clean / 1.51 at S5 versus 4.14 / 1.44). Notably, despite a 14× smaller backbone and no Open X-Embodiment pretraining, it matches or exceeds OpenVLA-OFT and OpenPi–0.5 on most corrupted settings—a strong claim that architectural inductive bias can partially substitute for scale. K-means clustering of projector outputs supports the mechanism: Fused IB-Adapter maintains coherent object-centric feature clusters under impulse noise, whereas standard MLP features diffuse into background-conflated representations.

Real-robot deployment

Experiments on the Astribot S1 dual-arm platform cover four tasks (pick-and-place, throw basketball, pour water, pack doll) under two digital corruptions (Gaussian noise, defocus blur) and two physical lens obstructions (oil, plastic shelter). StableVLA consistently shows the smallest performance drop of all tested methods—for example, an average drop of only −17.5 points on pick-and-place versus −30.1 for $\pi_{0.5}$ and −49.2 for VLA-Adapter—and achieves a 50% success rate on Pack Doll under corruption, exceeding both VLA-Adapter (20%) and OpenPi 0.5 (40%) despite having fewer parameters. Resilience to physical occlusions (oil, shelter) is particularly pronounced, which is significant because such disturbances are difficult to simulate exhaustively during training.

Ablations

Two ablations validate the design choices. Removing the MLP pathway (using IB-Adapter alone) reduces corrupted-data performance by 3.1 points on LIBERO and drops CALVIN from 2.13 to 1.44 completed tasks, confirming that high-fidelity spatial details must be retained alongside denoised semantics. Replacing Sigmoid with Softmax is more damaging: −16.3 points on LIBERO corrupted data and a collapse from 2.13 to 0.46 on CALVIN, empirically supporting the independent Bernoulli latent-structure assumption underlying the gating design.

Limitations and open questions

Several caveats bear directly on the results. First, the SPD dropout rate is tuned per benchmark suite (e.g., 0.0 for LIBERO-Long, 0.3–0.4 elsewhere), so the method requires task-specific hyperparameter calibration rather than a single universal configuration. Second, the theoretical derivation rests on Gaussian assumptions and a shared-covariance simplification; the equivalence between IB optimization and the proposed attention is approximate rather than exact in the deployed setting. Third, corruption evaluation excludes Glass Blur on three of four LIBERO suites and CALVIN due to computational cost, so coverage is not fully uniform. Fourth, the real-robot comparison includes only two baselines and relatively small trial counts (10 trials per condition), leaving statistical strength limited. Finally, the paper demonstrates robustness primarily against common-corruption-style perturbations; whether the covariance-gating mechanism generalizes to other distribution shifts (e.g., lighting changes, viewpoint shifts, embodiment transfer) remains an open question, as does whether IB-Adapter transfers to VLA architectures beyond the VLA-Adapter framework.

Conclusion

StableVLA identifies the vision-to-LLM projector as a principal locus of visual fragility in VLA models and addresses it with a lightweight (<10M parameter), information-theoretically motivated adapter that filters noisy channels via covariance-based sigmoid gating while preserving fine-grained spatial detail through a fused dual-pathway design. Without any additional data or corruption-aware training, the approach yields large zero-shot robustness gains over its baseline (35.2% average in simulation, 31.7 percentage points on real pick-and-place) and achieves robustness competitive with substantially larger, heavily pretrained models. The results support the broader thesis that robust architectural inductive biases at the modality-alignment interface can complement—or partially substitute for—data scaling in embodied AI.