{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2606.26443", "kind": "paper", "url": "https://arxiv.org/abs/2606.26443"}, {"id": "gh:baiqi-li/watchact", "kind": "repo", "url": "https://github.com/Baiqi-Li/WatchAct"}, {"kind": "website", "url": "https://baiqi-li.github.io/watchact_page/"}], "authors": ["Baiqi Li", "Ce Zhang", "Yu Fang", "Yue Yang", "Shangzhe Li", "Mingyu Ding", "Gedas Bertasius"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2606.26443", "source_url": "https://arxiv.org/abs/2606.26443"}, "categories": ["Robot manipulation", "Robotics", "Robotics & Embodied AI"], "description": {"en": "WatchAct evaluates robot manipulation from observed human behavior. Each instance pairs a human-action video and a language instruction with an aligned simulator scene and an executable LIBERO task. It covers 3,000 long-horizon instances across 14 tasks in four capability domains: Event Grounding, Procedural Reasoning, Implicit Intent Inference, and Episodic Reasoning. The evaluation protocol separately measures video-to-plan reasoning, policy execution under oracle plans, and full task completion, in simulation and on a Franka Research 3 robot."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2606.26443", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2606.26443", "title": "WatchAct"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["University of North Carolina at Chapel Hill"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2606.26443", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2606.26443", "source_url": "https://openalex.org/W7166031072"}, "key": "claire-radar:2606.26443", "modality": null, "name": "WatchAct", "openness": {"code_license": null, "data_license": null, "evidence": [], "status": "unknown"}, "provenance": {"confidence": "High", "crawl_bundle": "claire_radar_2026-10-14", "crawled_at": "2026-10-14T09:44:51Z", "data_status": "primary-source-indexed", "display_eligible": "unknown", "origin_record_id": "bm_watchact_107de5f7", "origin_source": "arxiv", "recognition_confidence": "1.0", "record_sha256": "e394da7038fdfa53fc22c43b826f4888aafa69be31e044742317636147c85095", "relation": "introduces", "review_state": "unreviewed", "source_url": "https://arxiv.org/abs/2606.26443"}, "publisher": null, "released": "2026-06-24", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2606.26443", "source_url": "https://arxiv.org/abs/2606.26443"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2606-26443", "source": "claire_radar", "source_benchmark_id": "2606.26443", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "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/Baiqi-Li/WatchAct", "githubScope": "benchmark_repo", "githubStars": 4, "hfDailySubmittedAt": null, "hfDataset": null, "hfDatasetDownloads": null, "hfDatasetLikes": null, "hfPaperUpvotes": null, "hfPaperUrl": "https://huggingface.co/papers/2606.26443", "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": {"checkedAt": "2026-10-04T15:31:48.030916Z", "state": "unavailable"}}, "status": "fresh"}, "capabilities": ["Robot manipulation"], "confidence": "High", "construction": "Interactive Environment", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-08-20T15:22:10.044731Z", "inputHash": "19fa056d94a3b2468d06aa862d3835c64ee43c1fa5bb28ac89328e2ec606faa3", "model": "deepseek-v4-flash"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "WatchAct evaluates robot manipulation from observed human behavior. Each instance pairs a human-action video and a language instruction with an aligned simulator scene and an executable LIBERO task. It covers 3,000 long-horizon instances across 14 tasks in four capability domains: Event Grounding, Procedural Reasoning, Implicit Intent Inference, and Episodic Reasoning. The evaluation protocol separately measures video-to-plan reasoning, policy execution under oracle plans, and full task completion, in simulation and on a Franka Research 3 robot.", "domainCuration": {"method": "rules-v1", "state": "auto"}, "evidence": {"reasonCodes": ["exact coined title identity tied to benchmark evidence", "benchmark release stated in one sentence", "evaluation protocol evidence", "public artifact URL"], "snippet": "We introduce WatchAct, a benchmark for robot manipulation grounded in observed human behavior."}, "familyId": "bmf_84d3a4a518a6", "firstSeenAt": "2026-08-19", "githubIndex": {"candidateCount": 0, "indexedAt": "2026-09-17T04:40:34.111665+00:00", "repositories": [{"evidence": [{"basis": "linked-from-project", "linkedUrl": "https://github.com/Baiqi-Li/WatchAct", "url": "https://baiqi-li.github.io/watchact_page/"}], "scope": "benchmark_repo", "url": "https://github.com/Baiqi-Li/WatchAct", "verification": "source-linked"}], "status": "linked"}, "heat": null, "id": "bm_watchact_107de5f7", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": ["Robotics"], "links": {"code": "https://github.com/Baiqi-Li/WatchAct", "data": null, "hfPaper": "https://huggingface.co/papers/2606.26443", "paper": "https://arxiv.org/abs/2606.26443", "pdf": "https://arxiv.org/pdf/2606.26443", "project": "https://baiqi-li.github.io/watchact_page/", "report": "https://arxiv.org/abs/2606.26443"}, "metrics": [], "motivation": "A robot working alongside people must reason about what they have done, in what order, and with what intent.", "name": "WatchAct", "oneLine": "WatchAct evaluates robot manipulation from observed human behavior. Each instance pairs a human-action video and a language instruction with an aligned simulator scene and an executable LIBERO task. It covers 3,000 long-horizon instances across 14 tasks in four capability domains: Event Grounding, Procedural Reasoning, Implicit Intent Inference, and Episodic Reasoning. The evaluation protocol separately measures video-to-plan reasoning, policy execution under oracle plans, and full task completion, in simulation and on a Franka Research 3 robot.", "paperTitle": "WatchAct: A Benchmark for Behavior-Grounded Robot Manipulation", "primaryDomain": "Robotics & Embodied AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2606.26443", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-10-14T01:46:58.058199Z"}, "publications": [], "ranking": {}, "readiness": "Runnable", "recognitionConfidence": 1.0, "relation": "introduces", "releasedAt": "2026-06-24", "source": {"authors": ["Baiqi Li", "Ce Zhang", "Yu Fang", "Yue Yang", "Shangzhe Li", "Mingyu Ding", "Gedas Bertasius"], "categories": ["cs.RO", "cs.AI", "cs.CV"], "id": "2606.26443", "title": "WatchAct: A Benchmark for Behavior-Grounded Robot Manipulation", "type": "arxiv", "url": "https://arxiv.org/abs/2606.26443"}, "sourceUpdatedAt": "2026-06-24T23:13:56Z", "topics": ["Robotics"], "venueAttempts": [], "whyItMatters": "Existing manipulation benchmarks typically evaluate from a single current image, lacking grounding in observed human behavior. WatchAct fills this gap by assessing robots' ability to reason about events, procedures, intents, and scene changes from video, which is critical for real-world human-robot collaboration. It provides a disentangled evaluation to isolate reasoning and execution failures, offering practical value for developing and comparing systems."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
