{"record": {"aliases": ["SafeVLA-Bench: A Benchmark for the Success-Safety Gap in Vision-Language-Action Models"], "artifacts": [{"id": "arxiv:2606.00773", "kind": "paper", "url": "https://arxiv.org/abs/2606.00773"}, {"kind": "website", "url": "https://safevla.org"}], "categories": ["Multimodal", "Robotics & Embodied AI", "Safety", "Safety & Trustworthiness"], "description": {"en": "SafeVLA-Bench evaluates safety-relevant trajectory behavior in simulator-based vision-language-action tasks using task success, unsafe-success frequency, and violation severity."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2606.00773", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2606.00773", "title": "SafeVLA-Bench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "key": "claire-radar:2606.00773", "modality": null, "name": "SafeVLA-Bench", "openness": {"code_license": null, "data_license": null, "evidence": [], "status": "unknown"}, "provenance": {"admission_policy_version": "2026-09-09.1", "confidence": "Medium", "crawl_bundle": "claire_radar_2026-10-14", "crawled_at": "2026-09-25T04:40:22Z", "data_status": "primary-source-indexed", "display_eligible": "true", "origin_record_id": "bm_safevla-bench-a-benchmark-for-the-success-_223bfd7e", "origin_source": "arxiv", "recognition_confidence": "0.55", "record_sha256": "0c905a961c24d87255e95414ee405fe0da20dd5ca799d46690322d4b67fd6c24", "relation": "evaluates_only", "review_model": "codex-cli-default", "review_state": "ai-reviewed", "reviewed_at": "2026-09-24T19:48:50.790988Z", "source_url": "https://arxiv.org/abs/2606.00773"}, "publisher": null, "released": "2026-05-30", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2606.00773", "source_url": "https://arxiv.org/abs/2606.00773"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2606-00773", "source": "claire_radar", "source_benchmark_id": "2606.00773", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": ["SafeVLA-Bench: A Benchmark for the Success-Safety Gap in Vision-Language-Action Models"], "annotation": "Unknown", "applicationDomains": ["Robotics & Embodied AI"], "area": "Safety & Trustworthiness", "attentionForecast": {"confidence": "Low", "horizon": "7d", "reason": "The explicit safety metrics and applicability to established VLA simulators may attract near-term robotics evaluation interest.", "score": 3}, "capabilities": [], "confidence": "Medium", "construction": "Interactive Environment", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-09-24T19:48:50.790988Z", "inputHash": "7fce838a788fabffc943356ac189b3fdc88368f5d211120fa8957bb63a737124", "model": "codex-cli-default"}, "curation": {"admissionPolicyVersion": "2026-09-09.1", "canonicalNameEvidence": "SafeVLA-Bench: A Benchmark for the Success-Safety Gap in Vision-Language-Action Models", "canonicalNameSource": "paper_title", "decisionReason": "The source names the evaluation, defines two concrete rollout-level metrics, and supplies a public project link.", "evaluationEvidence": {"reuse": "Project page: https://safevla.org.", "scoring": "Succ-But-Unsafe (SBU), the fraction of rollouts that both succeed and violate safety, and Violation Severity Index (VSI), a bounded worst-violation depth score", "task": "a post-hoc safety-evaluation framework for existing simulator-based VLA benchmarks"}, "model": "codex-cli-default", "reviewedAt": "2026-09-24T19:48:50.790988Z", "state": "ai-reviewed"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "SafeVLA-Bench evaluates safety-relevant trajectory behavior in simulator-based vision-language-action tasks using task success, unsafe-success frequency, and violation severity.", "displayEligible": true, "domainCuration": {"method": "rules-v1", "state": "auto"}, "evaluationMode": "public_reusable", "evidence": {"reasonCodes": ["coined title prefix ending in Bench or Benchmark", "benchmark release stated in one sentence", "evaluation protocol evidence", "public artifact URL", "named entity is a model, method, framework, or dataset used in evaluation"], "snippet": "We present SafeVLA-Bench, a post-hoc safety-evaluation framework for existing simulator-based VLA benchmarks."}, "familyId": "bmf_61a90eaa92c8", "firstSeenAt": "2026-08-19", "heat": null, "id": "bm_safevla-bench-a-benchmark-for-the-success-_223bfd7e", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": ["Robotics"], "links": {"code": null, "data": null, "paper": "https://arxiv.org/abs/2606.00773", "pdf": "https://arxiv.org/pdf/2606.00773", "project": "https://safevla.org", "report": "https://arxiv.org/abs/2606.00773"}, "metrics": [], "motivation": "Vision-language-action (VLA) benchmarks measure whether a policy completes a requested manipulation task, but binary success can hide safety-relevant trajectory behavior: reaching the goal while applying excessive contact, disturbing bystander objects, destabilizing the held object, or entering robot self-contact.", "name": "SafeVLA-Bench", "oneLine": "SafeVLA-Bench evaluates safety-relevant trajectory behavior in simulator-based vision-language-action tasks using task success, unsafe-success frequency, and violation severity.", "paperTitle": "SafeVLA-Bench: A Benchmark for the Success-Safety Gap in Vision-Language-Action Models", "primaryDomain": "Robotics & Embodied AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2606.00773", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-08-19T11:05:13.395391Z"}, "publications": [], "readiness": "Paper only", "recognitionConfidence": 0.55, "relation": "evaluates_only", "releasedAt": "2026-05-30", "reviewLifecycle": {"state": "pending"}, "source": {"authors": ["Jialiang Fan", "Weizhe Xu", "Oleg Sokolsky", "Insup Lee", "Fanxin Kong"], "categories": ["cs.RO"], "id": "2606.00773", "title": "SafeVLA-Bench: A Benchmark for the Success-Safety Gap in Vision-Language-Action Models", "type": "arxiv", "url": "https://arxiv.org/abs/2606.00773"}, "sourceUpdatedAt": "2026-05-30T15:31:57Z", "topics": ["Multimodal", "Safety"], "venueAttempts": [], "whyItMatters": "It distinguishes successful task completion from executions that violate explicit safety requirements."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
