{"record": {"aliases": ["RoboSPA: Can VLA Models Go Beyond Simple Scenes and Short-Horizon Tasks?"], "artifacts": [{"id": "arxiv:2609.05324", "kind": "paper", "url": "https://arxiv.org/abs/2609.05324"}, {"id": "gh:fanzhenxuan/robospa", "kind": "repo", "url": "https://github.com/fanzhenxuan/RoboSPA"}], "authors": ["Zhenxuan Fan", "Bo Zhang", "Yutong Lin", "Yuqian Yuan", "Juekai Lin", "Liang Liang", "Zhuoyi Huang", "Wenqiao Zhang", "Juncheng Li", "Siliang Tang", "Jun Xiao", "Yueting Zhuang"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2609.05324", "source_url": "https://arxiv.org/abs/2609.05324"}, "categories": ["Reasoning", "Robot manipulation", "Robotics", "Robotics & Embodied AI"], "description": {"en": "RoboSPA is a large-scale robotic manipulation benchmark that evaluates Vision-Language-Action models on fine-grained spatial reasoning and long-horizon procedural planning. It includes 10 task categories and 56 base tasks, each instantiated at five difficulty levels to yield 280 variants, with 527K trajectories. Beyond binary success, it reports diagnostic metrics such as ONTA and Progress Score. Code and data are available."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2609.05324", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2609.05324", "title": "RoboSPA"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "key": "claire-radar:2609.05324", "modality": null, "name": "RoboSPA", "openness": {"code_license": null, "data_license": null, "evidence": [], "status": "unknown"}, "provenance": {"confidence": "Medium", "crawl_bundle": "claire_radar_2026-10-14", "crawled_at": "2026-10-14T09:44:51Z", "data_status": "primary-source-indexed", "display_eligible": "false", "origin_record_id": "bm_robospa-can-vla-models-go-beyond-simple-sc_717e3ffd", "origin_source": "arxiv", "recognition_confidence": "0.7", "record_sha256": "edaac96af5f3c36530112b06b01177340180383476e08a6f5bd578fd5726a141", "relation": "introduces", "review_model": "deepseek-v4-pro", "review_state": "ai-name-audit-deferred", "reviewed_at": "2026-09-08T00:56:29.478814Z", "source_url": "https://arxiv.org/abs/2609.05324"}, "publisher": null, "released": "2026-09-04", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2609.05324", "source_url": "https://arxiv.org/abs/2609.05324"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2609-05324", "source": "claire_radar", "source_benchmark_id": "2609.05324", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": ["RoboSPA: Can VLA Models Go Beyond Simple Scenes and Short-Horizon Tasks?"], "annotation": "Unknown", "applicationDomains": ["Robotics & Embodied AI"], "area": "Robotics & Embodied AI", "attention": {"asOf": "2026-10-05", "attemptedAt": "2026-10-04T15:31:48.030916Z", "githubRepo": "https://github.com/fanzhenxuan/RoboSPA", "githubScope": "benchmark_repo", "githubStars": 19, "hfDailySubmittedAt": "2026-09-09T00:00:00.000Z", "hfDataset": null, "hfDatasetDownloads": null, "hfDatasetLikes": null, "hfPaperUpvotes": 28, "hfPaperUrl": "https://huggingface.co/papers/2609.05324", "observedAt": "2026-10-04T15:31:48.030916Z", "signalStatus": {"githubStars": {"observedAt": "2026-10-04T15:31:48.030916Z", "state": "fresh"}, "hfDatasetDownloads": {"state": "not_applicable"}, "hfDatasetLikes": {"state": "not_applicable"}, "hfPaperUpvotes": {"observedAt": "2026-10-04T15:31:48.030916Z", "state": "fresh"}}, "status": "fresh"}, "attentionForecast": {"confidence": "Low", "horizon": "7d", "reason": "Robotic manipulation is a high-interest area in AI research, and the paper provides extensive public resources including a dataset, code, and project page, likely to attract rapid community attention.", "score": 70}, "capabilities": ["Reasoning", "Robot manipulation"], "confidence": "Medium", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-09-08T08:59:06.803818Z", "inputHash": "e7b19aac43d232b6cf3de489f9b618cc25f26890d95d0ed9c475e90372a37a30", "model": "deepseek-v4-flash"}, "curation": {"canonicalNameEvidence": "We introduce \\textbf{RoboSPA} (\\textbf{Robo}t \\textbf{S}patial-\\textbf{P}rocedural \\textbf{A}ssessment), a large-scale robotic manipulation dataset and benchmark for diagnosing embodied reasoning in VLA models.", "canonicalNameSource": "abstract", "decisionReason": "The provided repository states the code and dataset are not yet released; without public artifacts, the benchmark cannot be independently run or verified.", "model": "deepseek-v4-pro", "reviewedAt": "2026-09-08T00:56:29.478814Z", "state": "ai-name-audit-deferred"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "RoboSPA is a large-scale robotic manipulation benchmark that evaluates Vision-Language-Action models on fine-grained spatial reasoning and long-horizon procedural planning. It includes 10 task categories and 56 base tasks, each instantiated at five difficulty levels to yield 280 variants, with 527K trajectories. Beyond binary success, it reports diagnostic metrics such as ONTA and Progress Score. Code and data are available.", "displayEligible": false, "domainCuration": {"method": "rules-v1", "state": "auto"}, "evaluationMode": "public_reusable", "evidence": {"reasonCodes": ["benchmark term in abstract", "benchmark release stated in one sentence", "evaluation protocol evidence", "public artifact URL"], "snippet": "We introduce \\textbf{RoboSPA} (\\textbf{Robo}t \\textbf{S}patial-\\textbf{P}rocedural \\textbf{A}ssessment), a large-scale robotic manipulation dataset and benchmark for diagnosing embodied reasoning in VLA models."}, "familyId": "bmf_3660cbd2b97b", "firstSeenAt": "2026-09-07", "heat": null, "id": "bm_robospa-can-vla-models-go-beyond-simple-sc_717e3ffd", "indexedAt": "2026-09-07T06:22:02.605340Z", "industrySectors": ["Robotics"], "links": {"code": "https://github.com/fanzhenxuan/RoboSPA", "data": null, "paper": "https://arxiv.org/abs/2609.05324", "pdf": "https://arxiv.org/pdf/2609.05324", "project": null, "report": "https://arxiv.org/abs/2609.05324"}, "metrics": [], "motivation": "Vision-Language-Action (VLA) models have shown promising progress in language-conditioned robotic manipulation.", "name": "RoboSPA", "oneLine": "RoboSPA is a large-scale robotic manipulation benchmark that evaluates Vision-Language-Action models on fine-grained spatial reasoning and long-horizon procedural planning. It includes 10 task categories and 56 base tasks, each instantiated at five difficulty levels to yield 280 variants, with 527K trajectories. Beyond binary success, it reports diagnostic metrics such as ONTA and Progress Score. Code and data are available.", "paperTitle": "RoboSPA: Can VLA Models Go Beyond Simple Scenes and Short-Horizon Tasks?", "primaryDomain": "Robotics & Embodied AI", "publication": {"evidence": "Accepted at the EMNLP 2026 Main Conference", "evidenceLevel": "author-claim", "evidenceUrl": "https://arxiv.org/abs/2609.05324", "source": "arxiv-comments", "status": "acceptance_claimed", "venue": "EMNLP 2026 Main Conference", "verifiedAt": "2026-09-07T06:22:02.605340Z"}, "publications": [], "ranking": {"90d": {"confidence": "Medium", "coverage": 0.7, "growth": {"components": {"githubStars": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}, "hfDatasetDownloads": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}, "hfPaperUpvotes": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}}, "confidence": "Low", "coverage": 0.0, "elapsedDays": null, "historyDate": null, "rank": null, "score": null, "windowDays": 90}, "level": {"components": {"githubStars": {"normalization": "log-max", "normalized": 0.3998764068733198, "value": 19}, "hfDatasetDownloads": {"normalization": "log-max", "normalized": null, "value": null}, "hfPaperUpvotes": {"normalization": "log-max", "normalized": 0.5949830216993645, "value": 28}}, "confidence": "Medium", "coverage": 0.7, "rank": 97, "score": 46}, "method": "attention-ranking-v13", "rank": 97, "score": 46}}, "readiness": "Runnable", "recognitionConfidence": 0.7, "relation": "introduces", "releasedAt": "2026-09-04", "source": {"authors": ["Zhenxuan Fan", "Bo Zhang", "Yutong Lin", "Yuqian Yuan", "Juekai Lin", "Liang Liang", "Zhuoyi Huang", "Wenqiao Zhang", "Juncheng Li", "Siliang Tang", "Jun Xiao", "Yueting Zhuang"], "categories": ["cs.RO", "cs.AI", "cs.CV"], "id": "2609.05324", "title": "RoboSPA: Can VLA Models Go Beyond Simple Scenes and Short-Horizon Tasks?", "type": "arxiv", "url": "https://arxiv.org/abs/2609.05324"}, "sourceUpdatedAt": "2026-09-04T16:19:01Z", "topics": ["Robotics", "Reasoning"], "venueAttempts": [{"decisionRaw": "Accepted at the EMNLP 2026 Main Conference", "evidence": [{"level": "author-claim", "observedAt": "2026-09-07T06:22:02.605340Z", "rawValue": "Accepted at the EMNLP 2026 Main Conference", "sourceType": "arxiv-comments", "sourceUrl": "https://arxiv.org/abs/2609.05324"}], "reviewStatus": "accepted", "venueName": "EMNLP 2026 Main Conference"}], "whyItMatters": "Existing VLA benchmarks often test only simple, short-horizon tasks under predefined conditions. RoboSPA scales spatial ambiguity and procedural complexity to reveal weaknesses in current models, providing a diagnostic tool for more capable embodied agents."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
