{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2607.27180", "kind": "paper", "url": "https://arxiv.org/abs/2607.27180"}, {"kind": "website", "url": "https://human-claw.github.io/"}], "categories": ["Multimodal", "Robotics & Embodied AI"], "description": {"en": "HumanCLAW-Bench evaluates how VLMs make embodied decisions. Its 1,218 humanoid find-navigate-interact episodes span 41 indoor scenes, and low-level motor execution is factored out. Models are scored by success rate."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2607.27180", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2607.27180", "title": "HumanCLAW-Bench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "key": "claire-radar:2607.27180", "modality": null, "name": "HumanCLAW-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_humanclaw-bench_f35ba6b5", "origin_source": "arxiv", "recognition_confidence": "0.55", "record_sha256": "2170677682d00b0b8e5b5c4596ec60774b4b3550cf743d16f221b97873353748", "relation": "evaluates_only", "review_model": "claude-code-default", "review_state": "ai-reviewed", "reviewed_at": "2026-09-24T20:47:54.872757Z", "source_url": "https://arxiv.org/abs/2607.27180"}, "publisher": null, "released": "2026-07-29", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2607.27180", "source_url": "https://arxiv.org/abs/2607.27180"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2607-27180", "source": "claire_radar", "source_benchmark_id": "2607.27180", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Unknown", "applicationDomains": ["Robotics & Embodied AI"], "area": "Multimodal", "attentionForecast": {"confidence": "Medium", "horizon": "7d", "reason": "Embodied VLM evaluation with low headline scores and a public project page may attract moderate attention.", "score": 55}, "capabilities": [], "confidence": "Medium", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-09-24T20:47:54.872757Z", "inputHash": "e10743e35cd67d819fad13407668ef42b9b033c33910c164a594f2f969667a33", "model": "claude-code-default"}, "curation": {"admissionPolicyVersion": "2026-09-09.1", "canonicalNameEvidence": "Based on this framework, we build HumanCLAW-Bench", "canonicalNameSource": "abstract", "decisionReason": "The benchmark is formally named and has a defined episode set. Its scoring uses success rate, and the source supplies a public project page link.", "evaluationEvidence": {"reuse": "Project page: https://human-claw.github.io/", "scoring": "the best model reaches only a 16.8% success rate", "task": "1,218 long-horizon, egocentric find-navigate-interact episodes across 41 indoor scenes"}, "model": "claude-code-default", "reviewedAt": "2026-09-24T20:47:54.872757Z", "state": "ai-reviewed"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "HumanCLAW-Bench evaluates how VLMs make embodied decisions. Its 1,218 humanoid find-navigate-interact episodes span 41 indoor scenes, and low-level motor execution is factored out. Models are scored by success rate.", "displayEligible": true, "domainCuration": {"method": "rules-v1", "state": "auto"}, "evaluationMode": "public_reusable", "evidence": {"reasonCodes": ["exact named benchmark artifact released in abstract", "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": "Based on this framework, we build HumanCLAW-Bench: 1,218 long-horizon, egocentric find-navigate-interact episodes across 41 indoor scenes."}, "familyId": "bmf_867ccb273de2", "firstSeenAt": "2026-08-19", "heat": null, "id": "bm_humanclaw-bench_f35ba6b5", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": ["Robotics"], "links": {"code": null, "data": null, "paper": "https://arxiv.org/abs/2607.27180", "pdf": "https://arxiv.org/pdf/2607.27180", "project": "https://human-claw.github.io/", "report": "https://arxiv.org/abs/2607.27180"}, "metrics": [], "motivation": "Evaluating whether a vision-language model (VLM) can act through a physical body is challenging.", "name": "HumanCLAW-Bench", "oneLine": "HumanCLAW-Bench evaluates how VLMs make embodied decisions. Its 1,218 humanoid find-navigate-interact episodes span 41 indoor scenes, and low-level motor execution is factored out. Models are scored by success rate.", "paperTitle": "HumanCLAW: Can Vision-Language Models Act Through a Body?", "primaryDomain": "Robotics & Embodied AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2607.27180", "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": "2026-07-29", "reviewLifecycle": {"state": "pending"}, "source": {"authors": ["Li Siyao", "Jiawei Gu", "Shuai Liu", "Kairui Hu", "Zekun Li", "Linjie Li", "Chengcheng Tang", "Po-Chen Wu", "Ivan Shugurov", "Lingni Ma", "Michael Zollhoefer", "Sizhe An", "Abhay Mittal", "Amy Zhao", "Ranjay Krishna", "Manling Li", "Ziwei Liu and Chuan Guo"], "categories": ["cs.CV", "cs.RO"], "id": "2607.27180", "title": "HumanCLAW: Can Vision-Language Models Act Through a Body?", "type": "arxiv", "url": "https://arxiv.org/abs/2607.27180"}, "sourceUpdatedAt": "2026-08-03T21:59:13Z", "topics": ["Multimodal"], "venueAttempts": [], "whyItMatters": "The benchmark separates a VLM's action choices from motor-control failures, so it isolates embodied decision-making. The paper reports that the nine VLMs tested reach at most 16.8% success."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
