Safety certification of Vision-Language-Action (VLA) driving planners under ISO 21448 (Safety of the Intended Functionality or SOTIF) requires an Operational Design Domain (ODD) specification that answers two complementary questions: when does the planner start to fail, and how severely does it fail once it does, according to a new arXiv paper titled When and How Severely: Scenario-Specific Safety Envelopes for Driving VLAs (arXiv:2606.14238). The research evaluates Alpamayo R1, a 10B-parameter open-weight driving VLA, on 15,968 (clip, attack) pairs and reveals that a single aggregate value per hazard is insufficient.
Key Findings: Conservative-Aggregate Gap and Severity Bands
The study identifies a conservative-aggregate gap: an aggregate safe threshold of σ ≤ 50 under a 15% average displacement error (ADE) budget masks well-sampled scenarios that tolerate the top of the tested grid (σ = 70). Using a Gaussian Mixture Model (GMM) on the changed-explanation subset, the researchers identify six discrete severity bands (BIC-optimal k=6), meaning two perturbation conditions with the same mean error can differ materially in their share of high-severity failures (C4/C5).
Joining the two analyses on the same corpus surfaces a finding neither yields in isolation: the scenarios with the loosest noise thresholds are not those with the lowest high-severity rate. Specifically, STOP_SIGNAL concentrates roughly 4× the C4/C5 share of LANE_KEEPING despite tolerating a larger σ.
| Scenario | Noise Threshold (σ) | High-Severity (C4/C5) Share |
|---|---|---|
| STOP_SIGNAL | Higher (σ up to 70) | ~4× that of LANE_KEEPING |
| LANE_KEEPING | Lower | Lower |
Implications for SOTIF Certification
The paper argues that a deployable SOTIF ODD specification for driving VLAs therefore requires a two-dimensional safety envelope, not a single aggregate value per hazard. This approach separates the two questions: when (noise tolerance) and how severely (high-severity failure rate), providing a more nuanced certification framework.
Methodology and Model
The evaluation used Alpamayo R1, a 10B-parameter open-weight driving VLA model. The researchers tested it on 15,968 (clip, attack) pairs, varying perturbation noise σ from low to high (grid up to σ=70). The GMM on the changed-explanation subset yielded six severity bands (C1–C6, with C4/C5 being high-severity). The study did not specify the exact attack types or dataset clips.
Background on the Standard
ISO 21448 (SOTIF) covers safety of the intended functionality for automated driving systems, focusing on hazards arising from functional insufficiencies or foreseeable misuse. The paper extends this to VLA planners, which combine vision, language, and action to drive autonomously.
The researchers—Priyadershi, Abhinaw, Frtunikj, and Jelena—published the paper on arXiv on June 12, 2026. The work highlights that scenario-specific safety envelopes are critical as VLA models grow in size and deployment scope.
For enterprise technology leaders evaluating autonomous driving systems for logistics or mobility, this study underscores the need for certification frameworks that account for scenario-specific failure modes rather than relying on aggregate metrics alone.
